Defining, Exploring, and Simulating

the Digital Twin Metaverses

Michael Grieves and Edward Y. Hua

Abstract This chapter presents a brief introduction and history of simulation, Digital
Twins and their types and replica twins, and the origin of “metaverse”. Underlying all
these technologies is the premise that information generated by these technologies
is a replacement for the wasted physical resources in human goal-oriented tasks.
The chapter then provides the characteristics of a Digital Twin-oriented metaverse.
It applies the characteristics to the different DT types. It concludes by discussing the
evolution of Digital Twins in replication and prediction that will see Front Running
Simulation as our crystal ball into the future. AI is predicted to play a major role in
making this evolution possible as an assistance to humans but not a replacement.

Keywords Digital twin · Physical twin · Replica twin · Metaverse · Simulation ·

Front running simulation

1 Introduction

Digital Twins (DTs) are a twenty-first-century concept that has enjoyed an expo-
nential growth of interest over the last decade. DTs originated as the underlying
component of another twenty-first-century concept, Product Lifecyle Management
(PLM). PLM represented a change from a functional-centric approach where each
function, engineering, manufacturing, operations, and support had siloed its data
and information to a product-centric approach where every function populated and
consumed from a common source.
That common source needed to be DTs. DTs took advantage of the exponen-
tial increases in information technology to implement this lifecycle-based, product-
centric representations of physical products and artifacts. This DT model was

M. Grieves
Digital Twin Institute, Cocoa Beach, FL, USA
e-mail: [email protected]
E. Y. Hua (B)
The MITRE Corporation, McLean, VA, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 1

M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_1
2 M. Grieves and E. Y. Hua

originally proposed for automotive and aerospace. However, DT use has been
proposed for almost any physical product or artifact that exists in the physical
world, both tangible and intangible. The chapters in this book include DTs for
industrial and manufacturing [1–4], nuclear reactors [5], health care [6], airports
[7], semiconductors [8], power generation [9], and batteries [10].
Humans have always had physical representations of different scales and fidelity
of products and artifacts in the form of Replica Twins as described later. DTs are a
digital representation of counterparts that exist or are intended to exist in the physical
world. Unlike atom-based objects where all or none of the objects exist, the bit-based
data and information of DTs allow for granularity of representation. The fidelity of
that granularity and the intervals needed to synchronize the DT and its physical
counterpart will depend on the use cases that create value. In order to understand
how value is created, it is important to understand what information does and the role
information plays with the different types of DTs throughout the product lifecycle.
DTs have gotten a significant amount of attention from both academics and
industry. However, equally important is a digital representation of the physical envi-
ronment that the DTs are or will exist and operate in. While it is necessary and
useful to have the data and information about the DT and its physical counterpart,
it is equally important to understand the forces of the environment surrounding and
affecting the DT and its physical counterpart. This means the creation of digital
environments or digital spaces1 that multiple DTs from different sources can inter-
operate in. These are what we are calling Digital Twin Metaverses that this chapter
will describe. This represents the next wave of evolution for DTs.
While defining and exploring DTs and DT Metaverses are important, this evolution
will feature the important capability of simulation. Simulation allows us to predict
and anticipate the future, at least probabilistically. We have had simulations for as
long as humans have been in existence and could think. What is novel is that we now
have the technology to simulate outside of human minds with DTs. This chapter will
set the stage for advancing this evolution.

2 Simulation

In the computer age, we technologists think that we have invented simulation.

However, humans have been doing simulations since the beginning of their existence.
Simulation is defined as one process that imitates another process [11]. It’s impor-
tant to note that processes are, by the very nature, time evolved. By that definition,
simulation is a foundational aspect of human thinking.

1We will use virtual and digital as synonymous here. There actually is no true virtual space. The
virtual representation is always instantiated in atom-based physical material. In humans, it’s in the
carbon-based matter of the brain. In digital computers, it’s in the silicon-based matter of digital
processor and memory components.
Defining, Exploring, and Simulating the Digital Twin Metaverses 3

2.1 Simulation Through Human History

Since man began to think, he has performed simulations. Man has used simulations
from the beginning of human existence for a wide variety of tasks. Simulation has
been used for planning, assessments, training, scenario generation, risk assessment,
experimentation, and even entertainment.
Take the example of prehistoric man hunting game. Prehistoric man ran through
different scenarios in his mind of what he, his fellow hunters, and the animal would
do. For example, the hunter mentally simulated what plan would predict running a
mammoth off a cliff. He ran through various simulations of which hunters needed to
be where, what actions they needed to take, and what reaction the animal would have.
He then selected the simulation that he believed would have the highest probability
of succeeding.
He shared that simulation with his fellow hunters by tracing his plan in the dirt
with a stick. His fellow hunters watched this simulation unfold over time as the hunter
traced the stick in the dirt, showing the movement of the hunters and their intended
prey. It was crude. It was primitive. It often didn’t predict the intended outcomes.
However, it was simulation.
The history of the military is intertwined with simulations. Soldiers throughout
history were trained in simulation exercises. Tzu [12] writes of convincing an emperor
of his ability to train troops by doing a simulation using the emperor’s harem. The
D-Day invasion of Normandy Beach was planned via simulation in a Scottish harbor
that had the characteristics of Normandy Beach [13].
Over history, humans developed more stylized simulations. Early Greek plays
were simulations of what would happen over time when certain events took place.
Later, in the Middle Ages, written stories were simulations [14]. In the 1900s, movies
came about and provided much richer simulations that could be shared by many more
people.
The arrival of computers in the last half of the twentieth century advanced the rigor
and robustness of simulations. These simulations were mathematically oriented and
could be quite complex in terms of calculations. However, these simulations with
applications like GPSS were mathematically abstract, and visualization was limited
to reams of numbers on paper output [15].
Fast forward to the twenty-first century. We can do very rich detailed simulations
that provide photo realistic visualizations of the simulation of physical objects in their
environments. These simulations reasonably mirror the changes of their physical
counterparts when subject to the same forces.

2.2 Simulation Prediction—Causation and Correlation

As noted above, simulations are about predicting possible future outcomes. There
are two mechanisms that simulation uses: causation and correlation. From a system’s
4 M. Grieves and E. Y. Hua

view of causation, we have inputs. We know those inputs cause specific things to
happen in our system. We then get outputs. It is deterministic.
In the correlation model, we have inputs. We don’t know deterministically what
happens in the system, but we get outputs. Even though we don’t know what happens
within the system, we do know that there is a relationship between the inputs and the
outputs. Varying those inputs will result in a varying of the outputs that maintain a
correlating relationship. The correlations may be very strong, close to 100% or very
weak, close to 0%. We can use strong correlations and ignore weak correlations.
We may have a correlation of 100% or close to that, but we may still not know why
or how the inputs result in the outputs. These correlation models can be useful even
if we have no idea why one input variable would correlate with an outcome variable
[16]. Note that we are very uncomfortable and should be questioning correlations
where we cannot theorize some relationship between correlated inputs and outputs.
Simulations can use both methods, causation and correlation. Simulation driven by
causation will use formula or algorithms to take inputs and derive outputs. Simulation
driven by correlation will take data that they have from previous input/output data
and apply that to the input data that they are attempting to determine outputs for.
Causation will give deterministic outputs, while correlations will give proba-
bilistic outcomes. However, it is important to note that for systems of high complexity,
there may be other unknown causal variables that will affect the outputs. We may
be unaware of these causal variables or that what we believe are causal variables are
simply correlated with this unknown causal variable.
Many of the computer-based simulations in the past were driven by formulas or
algorithms. The complexity of these simulations was such that this was sufficient for
the outputs that were needed. The simplest version of this is y = f (t). This could be
an object moving in a straight line with a constant velocity. The variable “y” is the
location at any time “t” produced by the function. While this is not usually thought
of as a “simulation”, it meets the requirements of the time-evolved definition. These
functions or formulae can be very complicated but are causally deterministic.
However, as we get into more and more complex systems with numerous variables,
some hidden, we will need to use correlations from big data that we are collecting
from products that are in operation. As we collect more and more data from products
in the field, Bayesian-based probability models assisted by Artificial Intelligence
(AI) become more and more useful [17].
Using causation when we have it is highly preferred. However, it needs to be
remembered that causal models are conditional and, for complex systems, may not
reflect or predict accurately or capture all causal relationships [18]. Probabilistic
correlation models can be highly effective in predicting overall successful outcomes.
We can also use a hybrid approach which combines using both causal and big data
correlations [19].
Going back to our prehistoric hunter, if he was successful 50% of the time in
running a mammoth off a cliff but not knowing exactly why it worked, he and his
tribe were well fed.
Defining, Exploring, and Simulating the Digital Twin Metaverses 5

2.3 Front Running Simulation (FRS)

FRS is a specific form of simulation. FRS was introduced by Grieves [20] and is
shown in Fig. 2. FRS is a simulation that predicts future states using assumptions
based on physics and/or data. However, instead of the initial conditions of the simu-
lation being arbitrarily set, the initial conditions are taken from the state of actual
conditions in the physical environment. At every new time zero (t0 ), FRS simulates
future states and attaches probabilities to those future states. The future states that
are the concern of FRS are states of adverse events. Adverse events are events that
waste resources hindering or preventing us from completing our task goals.
There are two versions of FRS. The first is FRS using inputs from only the physical
product itself to predict future states. The second is that FRS uses inputs from the
physical product itself and the environment to predict future states. FRS acts as
crystal ball into the future.
The specific conditions of FRS are:
. A simulation that contains behavioral assumptions of a corresponding physical
product’s future states based on physics causality and/or data probabilities, usually
Bayesian based.
. The initial conditions of each simulation at t0 are taken from the current state of
an object in the physical world and, optionally, the environment that surrounds it.

2.4 Value of Simulation

Humans and non-human life, which for wont of a better term we will call “nature”,
have two different approaches to existence. Nature tries all possible combinations
and lets the environment select the winners. Nature can do this because its only goal is
survival of the fittest, and it has effectively an unlimited time horizon and resources.
Nature also does not care about individual living organisms.
Humans, on the other hand, do care about individual living organisms, especially
their own. Humans do not have unlimited time horizons and resources. Humans also
have other goals besides survival. The human approach is to be task goal-oriented
and to accomplish that task using the minimum of physical resources, time, energy,
and materials.
Given that Is the case, the human approach means that for a goal-oriented
task, humans employ sophisticated thinking capabilities utilizing data, information,
knowledge, and wisdom (DIKW) [21]. This relies on the fact that the expended
physical resources used to perform the goal-oriented task can be divided up into
two categories. These two categories are shown in the left bar in Fig. 1 [22].2 The
lower part of the bar is the minimum amount of resources that is ideally required to

2The figure in the book had a third category, Execution Inefficiencies, where we know what we
need to do to eliminate the waste but don’t have the technical capabilities yet to do so. That category
was dropped in later versions in the interest of simplification.
6 M. Grieves and E. Y. Hua

complete the task. Everything above that is in the second category of wasted phys-
ical resources. We can apply a cost function to this in order to value the time, energy,
material of the physical resources needed to accomplish the task (Fig. 2).
The right-handed bar shows the role of information in task accomplishment. The
amount of physical resources necessary to accomplish the task in the most optimal
fashion stays the same. Information cannot replace the minimum amount of physical
resources necessary to complete the task. However, information can replace the infor-
mation inefficiencies or wasted resources (C w (x)) over and above that. For purposes
of illustration, this figure shows information as replacing all the wasted resources,
but for human endeavors, this will usually not be possible.

Fig. 1 Front Running Simulation (FRS)

Fig. 2 Information as time, energy, material trade-off

Defining, Exploring, and Simulating the Digital Twin Metaverses 7

The issue with this costing function is that the cost of generating information
from data3 does not come in a unit of measure, like physical resources. We can use
time and physical resources needed to develop this information as a proxy for a unit
of measure. In the past, that proxy has always been human resources and physical
materials. Today, the proxy commonly consists of computer hardware, software, and
energy in addition to human resources. As represented in the formula, the assumption
under which this model holds true is that the cost of developing this information is
less than the cost of wasting resources over all the times the task is performed.
Simulation is a method of developing information. Humans have no interest,
let alone the resources, in trying all possible combinations and letting the environment
dictate which one is successful. Humans want to simulate the possible ways of
obtaining their task-oriented goals and then perform the task using the method that
minimizes their scarce physical resources of time, energy, and material.
Throughout all of human existence, until very recently, those simulations involved
only the computational capabilities of human minds. The development and rapid
advancement of digital computing bring a quantum leap in simulation value
for human goal-oriented tasks. The more accurately predictive and cheaper that
their simulations are, the greater value humans will obtain in completing their
goal-oriented tasks while minimizing the use of physical resources.

3 Digital Twins

The concept of twins is ontological. We constantly categorize and compare things to

see if they are similar, so we can make decisions about how we should interact with
them. We expect to find similarities and regularities in our world. If we could not,
the world would be a very lethal place [23]. “Twin” is the regularity that something
is identical or nearly identical to something else, so that we can apply how one of
the twins looks and/or behaves to the other twin.
The only requirements for a “twin” are that it has two key attributes: duality and
strong similarity. That is there needs to be two of them, and they need to share signif-
icant attributes. There is no ontological or metaphorical requirement for timeline
simultaneity, as in human twins, scale similarity, or for the precedence of one type
of twin before another type of twin, e.g., a Physical Twin before a Digital Twin.

3 There is much confusion about what is data and information. We have a functional perspective.
Data is a fact or facts about reality and the input to create information: We collect data and process
it to create information. Information is a replacement or substitute of wasted physical resources:
We use information.
8 M. Grieves and E. Y. Hua

3.1 Replica Twins

There are physical objects that are “twins” in the sense that they are simply indepen-
dent duplicates. What we are interested in is 3D physical objects that are intended
to represent a specific physical object.4 Since the term “Physical Twin” (PT) refers
to the physical object in the dyad of Physical Twin–Digital Twin, we will call these
3D physical objects Replica Twins (RTs). We will use Replica Twins (RTs) in the
sense that there is a unique physical object and a replica physical object that can be
at different scales and fidelities.
RTs, even if primitive, rudimentary, and abstractly shaped, have been used in
human endeavors for all of human existence. RTs have been used at all scales, from
small models to full-size replicas and all fidelities, from exact replicas to simple
representations. While barely three-dimensional, prehistoric man sketched in the
dirt with a stick a representation of a mammoth, the cliff that the mammoth needed
to be driven over, and the positions of his fellow hunters. While military sand tables
date from the 1800s, equivalents date back to ancient Greece military use and most
likely before.
Architecture has used RTs from earliest times all over the world. RT artifacts
have been discovered dating from at least 6000 BC. RT model making was prevalent
in ancient Greece. The making of RT models to represent actual physical buildings
existed throughout the world in all cultures [24].
The RTs were even dynamic and not simply static. Watch any movie about World
War II. It will generally feature a table with a geographical map that people will move
around representations of military and naval forces. As dispatches come in, people
move these representations into different geographical positions so that commanders
can assess and plan their next strategic and tactical moves.
But it wasn’t simply scale models that were replica Physical Twins. Full-scale
RT mockups have been created and used. Full-scale RTs have been used in military
preparation as long as military engagements have existed. As noted above, the D-Day
preparation included exercises on Scotland beaches that were the replica physical
“twin” of Normandy Beach.
As discussed in the simulation section above, we could argue that since very early
times plays, and then later, in the twentieth century, movies have used full-scale RTs in
a form of simulation. Plays and movies have created exact replicas of existing physical
environments and then “simulate” activity within those environments. For example,
an exact replica of the US White House Oval Office appears in an innumerable
number of movies and television shows.
For equipment and vehicles, RTs were used primarily in development. However,
RTs were also used to resolve issues with operational equipment and vehicles.
Airplane manufacturers used replica twins to recreate and troubleshoot reported

4Obviously, there have always been 2D representations of physical objects, such as sketches,
drawings, blueprints. However, we wouldn’t call them “replicas”, as humans must do much mental
work to visualize them even poorly as three dimensional. They are more accurately described as
abstractions.
Defining, Exploring, and Simulating the Digital Twin Metaverses 9

problems with their airplanes in the 1930s. When problems were reported with auto-
mobiles, it was standard operating procedure for the engineer working on the problem
to find an identically configured automobile to try and recreate the problem.
As we developed electronics, we could make these RTs dynamic on their own.
The company one of the authors worked for in the 1970s, Lear Siegler Corporation,
had an F-16 flight simulator in its Grand Rapids Instrument Division. It was physical,
not digital. However, dynamic flight simulators date back to the 1930s with the Link
simulator [25]. There have been dynamic replica Physical Twins of nuclear reactor
control rooms for training and emergency exercises for over 50 years [26].
The Apollo program is often cited as the first use of Digital Twins. That myth is
still being perpetuated today.5 The common reference is to the Apollo 13 mission,
where the myth is that its “digital twin” was used to bring the crew safely back home
after an almost catastrophic malfunction.6
The reality is that the digital capabilities of the most powerful computer main-
frames of the Apollo days were extremely limited compared to today. Main memory
of the most powerful mainframes of the era was in the 16 MB range. The Command
and Lunar Landing Modules had a miserly 2K of main memory. The extensive trou-
bleshooting on earth was done with a series of RT capsules that had no “digital”
aspects.
Replica twins have been in existence throughout humanity’s history and are still in
use today. Replica twins have been abstracted in such representations as dirt sketches
and sand tables. Replica twins have been realistic scale models such as buildings and
even cities. Full-scale replica twins have been used to prepare for and track military
engagements. Replica twins have existed dynamically as in the Link, F-16, nuclear
reactor control rooms, and the Apollo space capsule simulators. The advent and rapid
development of digital computers enabled the logical next step of moving “twins”
from physical replicas to digital ones.

3.2 The Rise of Digital Twins

Digital Twins are a twenty-first-century concept. While considered a possibility since

the early days of digital computers [27], it isn’t until the 2000s that a cohesive model
and concept were proposed.
There were two major capabilities that DTs were intended to have: replication and
prediction. Replication is the characteristic that the DT would possess all the data of

5 This can be independently verified by doing a search of academic papers using the keywords
“digital twin” and “Apollo”.
6 Apollo 13 might be the most amazing malfunction recovery story ever. One of the authors had the

privilege of meeting the Apollo 13 Commander, James Lovell, and hearing first-hand the amazing
story of two astronauts sitting on what was basically a couch in Apollo 13, lining up the earth’s
meridian vertically and horizontally perfectly on a reticule etched on Apollo 13’s window so that
they didn’t burn up or bounce off into space at re-entry. However, the “twin” involved in working
the problem on earth was a replica twin capsule simulator, not a digital twin.
10 M. Grieves and E. Y. Hua

its physical counterpart needed for use cases. At its ideal, any data that could be had
while in physical proximity of the physical counterpart could be obtained from its DT.
The characteristic of prediction is that the DTs would causally or probabilistically
predict the future states of their physical counterparts.
The origin of the Digital Twin model is well documented in a multitude of
academic papers and industry articles [28]. Figure 3 is the first version of the
Digital Twin model that was presented at a Society of Manufacturing Engineering
(SME) conference in Troy, Michigan in October 2002. The presentation was entitled
Completing the Cycle: Using PLM Information in the Sales and Service Functions
[29]. It was about using Digital Twins in the operational and support phase of the
product lifecycle when there was both the physical product and its Digital Twin.
The model in Fig. 4 was refined a little later that year to emphasize that prod-
ucts existed in real and virtual spaces. This version was for a meeting of industry
executives, automotive software providers, and academics from the University of
Michigan. The meeting was to explore setting up the Product Lifecycle Management
Development Consortium (PLM DC) at the University of Michigan. Because of the
automotive industry attendees, the focus was on different product lifecycle phases
than the SME conference, namely engineering and manufacturing.
Both presentations were about the new discipline that was being defined, Product
Lifecycle Management (PLM). As a result, the model did not even have a name, as it
was simply entitled “Underlying Premise of PLM”. It did describe that the model’s

Fig. 3 Original Digital Twin model

Defining, Exploring, and Simulating the Digital Twin Metaverses 11

Fig. 4 Original Digital Twin model version 2

purpose was “Information Mirroring”. Even though it was rudimentary, the model
contained the major elements of the Digital Twin that exist today.
On the left side were physical products in physical space. These are the Physical
Twins (PTs). On the right side were virtual products, which we now refer to as Digital
Twins, in virtual space that corresponded to the physical products. The third element
is that there were communications between the two spaces and products, with data
from physical space and products obtained from sensing and IoT devices populating
the virtual space and products, and data and information coming back from virtual
space and products to be used in the physical space.
These models also contained the sub-spaces as part of the virtual space, VS1 ,
VS2 , VS3 … VSn. The idea of virtual spaces was fairly new at the time, so this was
to emphasize the fact that while there was only one physical space, there could be
an unlimited number of virtual spaces.
This model highlighted that there were two main functions that it implemented:
replication and prediction. Replication is the characteristic that the DT would possess
all the data of its physical counterpart. The products in physical space were replicated
by the products in virtual space. At its ideal, any data that existed in the physical
product was replicated in its Digital Twin.
The characteristic of prediction is that the DT would causally or probabilistically
predict the future states of the physical counterparts. The subspaces were virtual
12 M. Grieves and E. Y. Hua

Fig. 5 Current Digital Twin model

areas where predictive simulations could be performed. The multiple subspaces were
indications that there was no limit as to the number of simulations that could be done.
The model was changed during work at NASA. The graphics are better, courtesy
of NASA, when the model was used for a Department of Defense (DoD) conference
[30]. In addition, the model was also simplified, as shown in Fig. 5. It was felt that
sub-virtual spaces unnecessarily complicated the model. However, as we will explain
later, the original model better represents a metaverse model.
While the model did not have a name originally, it did acquire some names shortly
thereafter. It was called the Mirrored Space Model first, and then shortly thereafter
that was changed to Information Mirroring Model. The Information Mirroring Model
name remained albeit somewhat obscurely until around the 2010 timeframe. At that
time John Vickers of NASA who was working with Grieves suggested the name
Digital Twin. The Digital Twin name was a replacement for the relatively strange
name that NASA was using, Virtual Digital Fleet Leader.
The Digital Twin name was mentioned in a footnote in Grieves’ book on PLM
[31], attributing the name to John Vickers. Grieves used the Digital Twin name in
one of the seminal and highly cited Digital Twin papers [32], which noted that the
Digital Twin name was going to be used for the model from then on. Later, in a short
but highly influential piece, Grieves wrote an article for the Economist Magazine in
the GE Lookahead section that was subsequently picked up by the World Economic
Forum [33]. The Digital Twin was explained to the general audience in that short
article.
2015 marks the beginning of an exponential growth in reference to Digital Twins
in academic papers, industry white papers, and websites. The uses of DTs were
initially proposed for aerospace and automotive. That has exploded to encompass a
huge swath of industries and disciplines: power generation, heavy machinery, smart
Defining, Exploring, and Simulating the Digital Twin Metaverses 13

building/cities, oil and gas, ports and airports, archeology, and healthcare, just to
name a few. Doing an internet search of “Digital Twins” in 2018 resulted in only one
million hits [34]. A search of 2022 results in over 17 million hits. The number of
academic papers on Digital Twins shows a similar exponential growth (2015—295
results, 2022—17,100 results).
As of today, the model in Fig. 5 is the accepted model of the Digital Twin. While
definition may vary and vary widely, images usually show Digital Twin representa-
tions that are fairly consistent in the representation of physical space and products,
digital space and products, and a two-way connection between them.
The commonly accepted Digital Twin Model that was introduced in 2002 and
simplified to the one as shown in Fig. 5 consists of three main components:
. The physical products (PT) in the real-world environment.
. The Digital Twins (DTs) in a digital environment.
. The two-way connection between the physical and virtual for data and informa-
tion.
The third component, the connection between the physical and digital has often
been referred to as the “digital thread”. The connection is a two-way communication
connection. Data from physical products and their environment is communicated to
the digital environment and populate their DT counterparts for collection, assessment,
and response (CAR) [20]. Data and/or information (if action to replace physical
resources is proposed) is now available to the physical environment.
The digital environment of the Digital Twin, referred to as the Digital Twin Envi-
ronment (DTE), requires that it has rules that are identical as possible to our physical
environment. We need to be assured that the behavior of the Digital Twin in the DTE
mirrors the behavior of its physical counterpart for the use cases we require.
Finally, it is important to remember that when we refer to Digital Twins in a general
way, we are implicitly including all three elements of the Digital Twin Model. We
are not simply referring to the digital object that represents a specific physical object.
While it would be more accurate to use the term, Digital Twin Model, we simply us
Digital Twin.

3.3 Types of Digital Twins

There are three types of Digital Twins: the Digital Twin Prototype (DTP), the Digital
Twin Instance (DTI), and the Digital Twin Aggregate (DTA).
The DTP originates at the creation phase of the product lifecycle. The DTP of a
product begins when the decision is made to develop a new product, and work begins
doing just that. The DTP consists of the data and information of the product’s physical
characteristic, proposed performance behaviors, and the manufacturing processes to
build it. The DTP should also include the necessary processes and practices to ensure
the product is fully supportable and maintainable in the field and to troubleshoot and
repair the product effectively and efficiently to keep it operational. As much of this
work as possible should take place virtually.
14 M. Grieves and E. Y. Hua

The DTI originates when individual physical products are manufactured. DTIs
are the as-builts of the individual products and are connected to their corresponding
physical product for the remainder of the physical product’s life and even exist
beyond that. The DTI implements replication. Much of the DTI can simply be linked
to the DTP. For example, the DTI can link back to DTP 3D model and only needs to
contain the offset of exact measurements to the designed geometrical dimensioning
and tolerance (GD&T).
Pre-production physical product versions that are called physical prototypes will
have a DTI since these are actual instantiations of the developmental period. These
DTIs should be put to the equivalent digital tests and evaluations as the physical
prototype itself. Comparing the digital results to the physical results will increase
the confidence on relying on digital testing when the product is moved to full rate
production. Where there are significant deviations, digital testing can be improved
to converge on producing equivalent results to physical testing, with the goal of
dramatically reducing and even eliminating physical tests, except for a final physical
validation.
The DTP will contain the manufacturing process, Bill of Process (BoP), and the
parameters associated with the BoP. The DTI will contain any variations that occurred
in actual production. For example, the DTP process may require heat treating within
a temperature range. The DTI will capture the temperature that actually occurred.
The data and information that is needed for the DTI will be driven by the use cases
of the organization. The digital testing described above will make it possible to test
digitally each DTI of its physical counterpart to enable a high level of confidence in
the future performance of each individual product [35].
Because the DTI remains connected to its physical counterpart for the rest of that
physical product’s life, it will also contain the data from its operation. The DTI will
contain sensor readings and fault indicators. Based on use cases, the DTI will contain
a history of performance of state changes and resulting outcomes.
The DTA is the aggregate of all DTIs. The DTA contains all the data and infor-
mation about all the products that have been produced. The DTA may or may not
be a separate information structure. Some of the DTA data may be processed and
stored separately for analysis and prediction purposes. Other DTA data may simply
be mined on an ad hoc basis.
The bigger the population of DTIs, the more data that will be available to improve
Bayesian-based predictions. The DTA, which consists of the DTIs of physical
systems, is subject to model bias for its predictions. However, there are mathematical
techniques available for bias identification in DTs [36]. The DTA will also be the
source for Artificial Intelligence and Machine Learning (AI/ML) to predict expected
performance.
In 2019, Grieves introduced the Intelligent Digital Twin (IDT) to explain the
role that AI would have in both assisting Digital Twins in their performance and in
dealing with the increasing system complexity and emergent behavior of products
themselves [37]. The view here was that AI was not a replacement for humans but an
augmentation for humans. IDT specifies four attributes for Intelligent Digital Twins
as active, online, goal-seeking, and anticipatory.
Defining, Exploring, and Simulating the Digital Twin Metaverses 15

The characteristic of anticipatory requires that the Intelligent Digital Twin can be
constantly running simulations to look ahead into the future for its PT. That obviously
means that FRS is a critical component of Intelligent Digital Twins.

3.4 Digital Twins and Simulations

As defined above, simulations are one process that imitates another process. In the
case of Digital Twins, we require that the “process” we imitate is our physical
universe. Our DT simulations need to have as perfect fidelity to the laws of our
universe as we need for our use cases. The characteristics of materials and forces of
our physical universe need to be imitated as closely as possible. Simulation is what
is needed to implement DT prediction.
The one exception to adherence to the laws of our universe is the cadence of time.
We are unconstrained by time [37]. In our physical universe, time is completely out
of our control. We cannot go back in time. The only way to go forward in time is to
wait for the next tick of the clock. Even then, we go forward only according to the
set time. We cannot slow time down nor speed it up.
In digital spaces, we are time unconstrained and can completely control time.7
We can run our simulations at any clock speed. We can computationally go years and
decades into the future. We can also slow down time. We can break down actions
that happen in split seconds in our physical universe into microseconds.
In digital space, we can even go back in time which we cannot do in physical
space. In the physical world, we employ forensic methods to attempt to determine
what happened in the past that resulted in the current present. In simulations with
deterministic rules, we can usually reverse the arrow of time in the digital world. If
we have been using traceability to capture state changes, we can simply step back
through the time frames.
However, another fundamental advantage is prediction, being able to advance the
clock to see what’s going to happen in the future. This is a crystal ball that sees into
the future. With Front Running Simulation (FRS) described previously, we have the
ability to do just that.
The assumption is that “simulation time and wall clock time can be kept in sync
using conservative and optimistic synchronization protocols” [38]. At every new
time t0 , we will be able to take the data from the physical world, i.e., replication,
and predict at least probabilistically what will happen in the digital world. This is a
tremendous opportunity to prevent the waste of physical resources by anticipating

7 I discovered this firsthand in the early 1970s. Even though I was only a sophomore in college, I
was a systems’ programmer for a computer timesharing company. We charged by the CPU second.
There was a meeting to discuss how to increase revenue. After listening to the staff provide ideas,
I simply said, “I can increase what a CPU second is”. That worked in increasing revenue until
some customers started running benchmarks and complained that their programs were taking more
elapsed CPU seconds and therefore were more expensive to run!
16 M. Grieves and E. Y. Hua

and correcting adverse events, especially ones involving the safety of individuals,
before adverse events can occur.

3.5 Digital Twins, Simulation, and Information

The purpose and value of DTs are that they provide information that can replace
wasted physical resources. One of the three elements of the Digital Twin Model in
Fig. 5 is the connection between Physical Space and its Physical Twins and Virtual
Space and its Digital Twins. Data is sent from the physical products and optionally
the physical environment to the digital DTIs. The information that is created and
housed in digital space is used in our physical space. This information is created
by processing the data coming from the physical space and, as noted above, by
performing simulations.
This information can take a couple of forms. It can be the result of humans doing
queries of DTs and creating information by using the result of the queries to take
action that will replace wasted resources.
For example, a certain model fuel pump of a helicopter is being recalled and
replaced. The traditional method would be to identify the location of every helicopter,
send a mechanic to inspect each one, and replace the fuel pumps in the helicopters
with the defective ones. With DTs, a query would be run on all the DTIs. Mechanics
would only be sent out to only those helicopters with defective fuel pumps. That
results in information replacing the wasted physical resources of mechanics’ time
and expenses for the helicopters with fuel pumps that need no replacing.
Information can also be created by routines that run on a constant basis in digital
space looking for specific sensor data patterns that, using physics and DTA data,
simulates and predicts adverse events and alerts humans to them. Humans have the
responsibility of deciding the actions to take to avoid wasted physical resources by
having an adverse event occur. This is a human-in-the-loop version.
The information can also be routines that run in digital space that specify actions
to take when they find certain conditions. These actions are coded as commands
and sent to the PTs directly, without human intervention. FRS is applicable in both
simulation-based situations. All are the result of data coming from the physical world
to digital world and data and information returning from digital world to the physical
world to replace wasted physical resources.

4 Metaverse

“Metaverse” is a portmanteau of “universe” and “meta”. “Universe” would

commonly be thought of as our one and only (“uni”) physical universe. “Meta”
is having aspects or capabilities that transcend or are beyond the ones our universe
possesses.
Defining, Exploring, and Simulating the Digital Twin Metaverses 17

Like Digital Twin, “Metaverse” is a twenty-first-century term. Although “meta-

verse” has its origin in the last decade of the twentieth century as described below,
“metaverse” only starts to be prominent in the last few years. The metaverse, virtual
spaces, virtual worlds, digital spaces, and digital worlds have been used fairly synony-
mously. All these terms refer to virtual representations of the physical world. As
explained in Product Lifecycle Management: Driving the Next Generation of Lean
Thinking [39], humans have had virtual spaces since the beginning of their existence.
Humans have always had the ability to virtually represent the physical world in
their own minds. However, those representations suffer from vagueness, imperma-
nence, and an inability to share with other humans. The result was that whenever we
developed a mental model or a new idea, it had to take immediate physical form in
some fashion: a sketch, a drawing, a blueprint, or a physical model.

4.1 Origins of Metaverse

The term, “metaverse” is almost universally attributed to originating in Neil

Stephenson’s book, Snow Crash [40]. Stephenson’s metaverse is a singular space
where an individual as an avatar interacts with other avatars in an immersive envi-
ronment. The immersive environment was implemented by a Virtual Reality (VR)
headset that the user wore.
This type of metaverse was almost exclusively about social interaction. This was
described in a contemporaneous definition of “virtual worlds”, which defined them
as, “A synchronous, persistent network of people, represented as avatars, facilitated
by networked computers” [41]. The virtual worlds and Stephenson’s metaverse were
synchronized with our physical world, could not be paused, and existed for the
interactions of people as avatars.
There was some correspondence with our physical world in that it had three
dimensions, and inanimate objects had persistence and impermeability. Avatars could
move about in this three-dimensional space. There was a monorail that traveled
the entire metaverse space and other vehicles were available, although how avatar
locomotion itself was accomplished was never really explained [42]. Like Linden’s
Second Life, users could acquire virtual real estate where they could construct virtual
buildings and other virtual artifacts [43].
Stephenson’s metaverse had its own laws that loosely, if at all, followed the laws
of our physical universe. For example, one can be driving a motorcycle at Mach 1
speed, run into a metaverse building, and the motorcycle and rider simply come to
a stop with no damage to anything.
In the years after Stephenson’s book and specifically within the last few years,
metaverse has become synonymous with almost any and all digital spaces. This
has led to the term “metaverse” applying to all digital spaces even if they are in
contradiction to each other.
There has been discussion of an “Industrial Metaverse” that is vaguely defined
as “the set of metaverse applications designed for industrial users” [44] [45]. The
18 M. Grieves and E. Y. Hua

descriptions on an Industrial Metaverse almost always focus on technologies that

users can exploit such as Augmented Reality and Virtual Reality. But there is a
lack of defining what is required for such a metaverse. This needs to be much more
concrete as we will discuss below.

4.2 Metaverse and Digital Twins

The original view of the metaverse is substantially different from what is required
for the purposes of Digital Twins. There is no reason that there cannot be all types of
metaverses with different rules governing them. The rules can be completely arbitrary
and have only a tenuous connection with our physical universe, as was the case for
the Stephenson metaverse.
However, we would contend that we can legitimately co-opt the term because our
metaverse required for DTs has more fidelity to the meaning of the words, “meta” and
“universe”. What we require is a digital fidelity to our physical universe with meta
capabilities. The metaverse that we require needs to have complete conformance
to the selected rules of our unitary, physical universe, although it can have meta
capabilities as described below.
In order to move work from the physical world into the digital world successfully,
the metaverse of Digital Twins needs to enforce all the laws of our physical universe
required by the supported use cases. However, we need a metaverse that is specifically
tailored to Digital Twins. From now on, we will refer to this type of metaverse as a
“DT Metaverse” (DTM).
As humans perceive it, our universe is a three-dimensional space populated by
objects. The DTM will have that same characteristic, a three-dimensional space
populated by DTs. This will allow humans to use their ontological understanding of
physical space to understand and operate seamlessly in the DTM.
Figure 6 is an evolutionary model of the DT that has been proposed previously
by Grieves [27]. The claim is that we are currently in Phase 2, the Conceptual Ad
Hoc phase. This is the current state of DT evolution.
In this phase, “the Digital Twin is an entity that we conceptually create from
disparate and even fragmented data sources. We use different existing sources to pull
data from. We start building correlations and even causations of data source inputs
to results. We build different simulation views and determine how well they map
to reality. We start to put manual processes in place to pull the data from different
sources, even if on an ad hoc basis, to create a Digital Twin view [27].
The next phase in the evolution of DTs is the creation of DT platforms. These
platforms are envisioned to support multi-users and multiple DTs. DT systems are
possible in Phase 2. However flexible Systems of Systems (SoS) where multiple DTs
of completely different functions enter and exit will require platforms that support
interoperability.
First will come platforms that support replication, Phase 3. Those platforms will
be followed by platforms for prediction, Phase 4. While not referred to as “DT
Defining, Exploring, and Simulating the Digital Twin Metaverses 19

Fig. 6 Digital twin evolution

Metaverses”, that is in essence what these platforms are by our definitions here.
These platforms now have a name, DT Metaverse. They are consistent with the
long-term vision of DTs.
The DT Metaverse has these key characteristics:
. There are multiple DT Metaverses.
. DT interoperability is a core requirement.
. All laws of the physical universe are implemented and enforced in simulations
for all inanimate objects.
. The DT Metaverse supports both replication and prediction.
. Multiple immersive participants as avatars is supported.
. Meta capabilities are allowed for human participants as avatars.
. Time can be synchronous or asynchronous with physical time depending on use
case and DT type.
. Cybersecurity is embedded in all aspects of the DT Metaverses.

4.2.1 There Are Multiple DT Metaverses

The limitation of our physical universe is that there is only one. If the Many Worlds
theory is correct and multiple physical universes exist [46], we will only ever have
access to the single physical universe this version of us inhabits. In the digital
universes, we can recreate our physical universe at some level. Since we have no
physical restrictions on the number of computer-based spaces, we have no restriction
or limitation of the number of digital metaverses that will be available to us.
20 M. Grieves and E. Y. Hua

As the original DT model proposed, we have effectively an unlimited N number

of sub-virtual spaces. Ironically, the original Digital Twin model in Figs. 3 and 4 is
a much more accurate model of the DT Metaverse than the simpler model in Fig. 5
that is evolved from it.
We expect that there will be many different metaverse platforms in the near future.
They will be differentiated by different use cases, different DT types, and different
phases of the product lifecycle. DT Metaverses can be on a spectrum of completely
private to completely public. While we are restricting our discussion to DTs of
tangible products, there is no reason that metaverses for intangible or process DTs
such as supply nets, manufacturing processes, or financial systems cannot be devel-
oped. If we can visualize the data and information in some symbolic fashion, we can
have a DT Metaverse for it.

4.2.2 DT Interoperability Is a Core Requirement

The concept of a DT Metaverse implies that there will be multiple DTs in it. Other-
wise, the DTM would be no different than what we have today, which is a Phase
2 Conceptual Ad Hoc programming space. The ability to easily insert a DT into a
digital environment to monitor its interaction and behavior with other DTs will greatly
enhance their value and usefulness. Interoperability also implies intra-operability as
this will also foster component modularity within DTs.
This means that the DT Metaverses are platforms. Platforms are hardware and
software infrastructure that provides underlying tools, services, and governance to
accomplish participative specific tasks and interactions [47]. Participative implies
that there is intended to be multiple users and multiple DTs. So, this means that the
platforms need to enable the interoperability of DTs.
There are numerous organizations working on mechanisms for interoperability.
A common perspective for interoperability is to produce standards and ontologies.
While this works in the physical world, it is much more difficult in the digital world
because of the much finer granularity of data and information. As a result, to address
this, we have a multitude of standards, which means that we really don’t have “a”
standard.
These other mechanisms that platform may employ include defining ad hoc
programming conventions, harmonization of programming conventions among soft-
ware providers, and a platform’s own middleware. A promising solution may be to
deploy AI. AIs may be able to explore the solution spaces between different DTs
and provide mapping and translations. As Fig. 7 illustrates, AI may provide both
intra-operability for components of a Digital Twin and interoperability for different
DTs in DTMs.
Depending on the DT type and use case, the DT Metaverse will be useful for
its ability to support immersion for multiple people assessing the serviceability of
a new product. It may also be useful for staging and operating multiple products in
simulation from different vendors in virtually commissioning a production line.
Defining, Exploring, and Simulating the Digital Twin Metaverses 21

Fig. 7 AI DTM interoperability at the system of systems level

4.2.3 The DT Metaverse Supports Both Replication and Prediction

There are two main characteristics that DTs need. Those characteristics are replica-
tion and prediction. Even if a product is being developed in digital space and has not
yet taken physical form, the environment that the new product is going to exist in
needs to replicate our physical universe for the use cases needed.
When we have DTIs of existing physical products, we need to replicate the data
from those PTs that we require for our use cases. The DT Metaverses supporting
this will be on a spectrum. On one end, there is simply the DTI in empty space
that can be simultaneously and instantaneously interrogated no matter where its PT
is in the physical world. On the other end of the spectrum, we can have DTIs and
the surrounding environment it currently resides in that an immersive avatar can
experience.
We need prediction both when we are developing the product and then when we
have a product that is in operation. We need to predict what forces that the product
produces will have on its own structure, its operation, and its surrounding environ-
ment. We also need to predict the impact that outside forces we’ll have on our product.
Prediction is done via time-evolved simulations and can be done based on hybrid
physics/data probabilities techniques. Again, the DTM will be on a spectrum from
simply having the DTI representation of PT itself to having the DTI representation
of PT and its current surrounding environment. This gives us the ability to be time
and space unconstrained with a probabilistic window into the future.
22 M. Grieves and E. Y. Hua

4.2.4 Laws of the Physical Universe Are Implemented and Enforced

in Simulations for Inanimate Objects

The primary requirement of the DT Metaverse is that inanimate objects will behave
as they would in our physical world. In DTM simulations, if, as described above,
a DT motorcycle traveling at Mach 1 slams into a stationary object like a building,
that will generate a tremendous amount of force as defined by F = M * V 2 . That
resultant force will deform/destroy the motorcycle, the building, and the surrounding
location.
Since the physical world is a tremendously complex place, it may not be feasible
to enforce all the physics of our universe. DT Metaverses will need to establish
the parameters of their adherence to our universe’s physical laws so that users will
understand what they can and cannot rely on. DT Metaverses will need to define the
use cases that they support.

4.2.5 Multiple Immersive Participants as Avatars Is Supported

The DT concept has always been intended that there be a social aspect in terms of
having many people collaborate in a shared environment [32]. However, it was orig-
inally intended that people be outside observers. When there is immersive capability
with people having avatars, we will need to carefully define and design the meta
capabilities.
What seems to differentiate metaverse from virtual environments is that meta-
verses enable immersion and host multiple participants. Immersion can be done first
person using avatars and Virtual Reality (VR) and Extended Reality (XR) capa-
bilities. The metaverse is a 3D geographic space that contains visual DTs that can
be examined and optionally operated by the avatars. The requirement for avatars is
important for knowing what participants are present in the metaverse and to know
who is operating the equipment, so multiple avatars don’t try and engage with DTs
simultaneously.
The DTMs will allow for multiple participants who have independence within the
metaverse. Participants are not constrained to a certain view but can move around
the metaverse independently of other participants.
This does not mean that it requires an immersive participant to use and obtain
value from the DT Metaverse. Authorized outside observers can have access to a
DT Metaverse and benefit from both replication and prediction data. Whether or
not observers can be active participants in a DT Metaverse will depend on the rules
employed by that DT Metaverse.
Defining, Exploring, and Simulating the Digital Twin Metaverses 23

4.2.6 Meta Capabilities Are Allowed for Human Participants

as Avatars

While it is foundational that our DT Metaverse follows the laws of our physical
universe, it does not preclude having meta capabilities for our participants of the DT
Metaverse. As described above, they can take advantage of the capabilities that simu-
lation and virtuality afford. Participants are unconstrained by location if so allowed,
meaning that they can “teleport” from one location to another. Solid structures can
become transparent so as to see inside. As long as the inanimate objects are subject
to the laws of our physical universe, human participants can have meta capabilities.
If the use case is that participant avatars are merely observers who simply want
to experience the contents of the DT Metaverse as would be seen in the physical
world such as the layout of a factory and its equipment, we can allow them meta
capabilities such as walking through solid objects. However, if the use case involves
people interacting with the DT Metaverse physical objects, their avatars need to be
subject to the physical rules like every other inanimate object. For use cases such as
physical factory walkthroughs or equipment training, avatars should not have meta
capabilities. That would defeat the purpose of these types of use cases.

4.2.7 Time Can Be Synchronous or Asynchronous with Physical Time

Depending on Use Case and DT Type

Another major difference with DT Metaverses from Stephenson’s metaverse is that

these DT Metaverses can be asynchronous and non-persistent with the physical
world. The original metaverses were synchronous and persistent. The time in the
metaverse was the exact same time as in the physical world and moved at exactly the
same cadence. The metaverse was persistent in that if people left the metaverse, the
metaverse continued without them.
A claim about metaverses is that persistence and synchronicity with the physical
world are a requirement for all metaverses [48]. This would be a requirement of a
DT Metaverse if the use case is real-time replicative and multiple PTs corresponding
to their DTIs exist in the physical world. However, the ability to internally control
time is a key advantage and capability of the DT Metaverse for some DT types and
use cases that involve prediction of future states.
As discussed above, depending on the DT type and use cases, time can be either
synchronous with physical time or asynchronous. However, time must be persistent in
either case. This means that time within a DTM must be the same for all entities inside.
If time speeds up or slows down, it must speed up or slow down for all inanimate
objects. Human participant avatars will simply be brought along as observers with
the time as a “fast forward” or simply with a discontinuous jump.
However, as noted above, if we are not time synchronous with the physical world,
we are time unconstrained. We cannot only go forward in time. We can reverse time’s
arrow and go backward. Again, there must be persistence for all DTs with human
avatars participating as described above.
24 M. Grieves and E. Y. Hua

4.2.8 Cybersecurity is Embedded in All Aspects of the DT Metaverse

It is not possible to do justice to cybersecurity as a chapter topic, except to highlight

its importance. A fortress approach with a remote physical location, impenetrable
walls and fences, and armed guards to secure physical products is feasible. A fortress
approach is not feasible for DTs and DTMs.
That makes cybersecurity the highest risk for the adoption of DT Metaverse capa-
bilities. The approach to DT Metaverse security needs to have three characteristics:
being embedded in all aspects, not just access, continuous monitoring of all partic-
ipants and interactions, and driven by extreme paranoia. If cybersecurity cannot be
guaranteed, DTMs cannot succeed.

5 Digital Twins, Simulation, and the DT Metaverse

The spaces below the DT Metaverse, indicated by VS1 , VS2 , VS3 … VSn, are virtual/
digital spaces, “sandboxes” is the term usually used, where different simulations of
our DTI could be performed. The initial conditions from the DT Metaverse would
all be the same, but other conditions and assumptions would be varied to understand
what different outputs would result. The main space and the sub-virtual spaces were
what was referred to above as the Digital Twin Environment (DTE) above.
In this section, we will put together the concepts we have described above.

5.1 DT Types and Use of Metaverses

As stated above, the DT Metaverse will not be a singular entity. There can and will
be different DT Metaverses, depending on where we are in the product lifecycle,
what DT types we are working with, and what use cases we need.
In terms of capabilities, as described above, the DT Metaverse must twin the
physical universe consistent with our physical laws. The DT Metaverse needs to
reflect what the physical world does. It needs to have persistent physical atom-based
material characteristics. This is called coherence that the physical world enforces
through atoms but needs to be specifically programmed into digital worlds [22]. For
example, if we lengthen a steel beam, it has to not only lengthen the beam, but it
must add physical weight.
It must have persistent forces active upon atom-based materials. If a force acts
on a particular material, it must do it consistently across all the materials in the DT
Metaverse. We cannot have forces acting one way on part of the DT and in a different
way on a different part of the DT.
We must have persistent time in each DT Metaverse. All the objects of our DT
Metaverse must have the same time basis. We cannot have some DTs speed up in
time and some DTs lag in time. There must be a persistent time basis in the DT
Metaverse. This requires Digital Twin interoperability. As we move different DTs
Defining, Exploring, and Simulating the Digital Twin Metaverses 25

into a DT Metaverse, they all have to participate as if they were physical objects in
the physical universe. Interoperability of different DTs is a major issue that we need
to address.

5.1.1 Digital Twin Prototype

The DT Metaverse of the DTP will need to support both replication and prediction.
In this case, since we are developing a new product, the replication will be of the
environment that our product is going to operate in. It will be a synthetic replication
that we will put our new product through its paces in. The primary focus of the DT
Metaverse in this lifecycle phase will be prediction of behavior over time.
We will want to predict the suitability for purpose and behavior performance of
our new product under a variety of conditions. We want to predict our product’s
performance throughout its entire lifecycle. As has been specified, we will want
to create the product virtually, test the product virtually, manufacture the product
virtually, and support the product virtually. Only when we get the product as perfected
as we possibly can do, we want to move around expensive atoms to build a physical
version.
The DT Metaverse will need to adhere to all the laws and rules implemented in
our physical universe for the use cases it supports. We want to test our products under
a range of conditions to see that they can measure up. We will want to put a new
airplane through years of missions that it will encounter. We want our pacemaker to
encounter a wide range of patient conditions and react with the required response
accordingly.
We want to manufacture a product on a virtual production line. What we want to
do here is reflect the movement of materials, machines, and people. We will explore
different Bill of Processes (BoPs) to determine the “leanest” version. We will look
at the ergonomics of the factory workers to ensure that there are no health and safety
issues. As a byproduct, we can do our virtual commissioning of new production
facilities during this phase.
In the support/operation phase, we will want to insert problems into our DTP and
see how we troubleshoot that particular issue. We want to disassemble and reassemble
to ensure that it will be easy and convenient in the field. We want to see how products
interface with support equipment such as refueling systems, to make sure that it has
been designed appropriately.
We want our DT Metaverse to support interoperability. This will allow us to take
components from different manufacturers and make sure that they interface with each
other. Interoperability also means that the physical laws are enforced on a coherent
basis. When we fill the fuel tank in the wings of an airplane, the plane fuel gauge
will register the increase in fuel, the wings will deform appropriately for the amount
of new mass they have in it, and the weight of the entire airplane will increase.
Humans will be able to participate via avatars. They will be able to walk around the
product, inspect areas of it, open doors and latches, and even operate the product. For
an airplane, they can get into the cockpit and fly the airplane through a mission. Using
26 M. Grieves and E. Y. Hua

VR and haptic equipment, they can bring in a refueling hose, connect it to the refueling
port, and feed fuel into the airplane. They can participate on the manufacturing line
and move parts from one place to another. They can operate and change settings on
the digital machines that are producing the parts.
These avatars can have meta capabilities, so they could jump from one place to
another or see through outside structures to see what’s underneath the surface cover-
ings. Avatar participants will have an indication as to whether time is synchronous
or asynchronous, so they can understand whether the simulation is being sped up or
slowed down.

5.1.2 Digital Twin Instance (DTI) and Digital Twin Aggregate (DTA)

We will consider the DTI and DTA at the same time, since the DTA is made up
of DTIs. When we move to looking at the DTI and DTA type, the DT Metaverse
requirement changes. The DT Metaverse is now driven by a Physical Twin. The PT
is at the mercy of our physical universe clock. PTs can only move forward in time
as fast as the clock ticks.
Even though there is a unique DTI for every PT object, there is no reason that the
DTI cannot participate in multiple DT Metaverses. There may be different use cases
that are involved that the different DT Metaverses are concerned with. What will be
true for replication use cases is that in each of the DTMs, the DTI in that DTM will
be identical at times that are synchronous to our physical universe to the same DTI
in other DT Metaverses.
The DT Metaverses for the DTI and DTA will support both replication and predic-
tion. At this phase in the lifecycle, there is actually a PT to replicate and the physical
environment that the PT lives in. Depending on the use cases, we need to populate
the DT Metaverse with both the PT data and the environmental data.
Because we have sub-virtual spaces, we can engage in prediction. This is where
FRS comes in the play. At every time zero, we can run a simulation in the sub-virtual
spaces to predict the future of individual products or even a group of products or prod-
ucts as a whole. We can use the Digital Twin Aggregates to provide Bayesian-based
probabilities. For purely deterministic situations, we can simply predict outcomes,
such as predicting that unless a ship reduces its speed immediately that it will be
unable to stop in time and will crash into its berth. In other cases, we may provide
probabilities of different outcomes.
As noted above, for real-time replicative use cases, DT Metaverses will require
that time will need to be synchronous and persistent. However, that will not preclude
use cases where it may be advantageous to drop out of real time in replication and
slow the happening of an event. This will give humans the ability to assess all that is
happening at a slower rate to make better decisions. Better decisions mean reducing
wasted physical resources to accomplish the goal by selecting the best future actions
to take.
Defining, Exploring, and Simulating the Digital Twin Metaverses 27

Obviously, our DT Metaverse will need to follow the laws and rules of our physical
world. While prediction of the PT itself is important, the impact that this PT will
have on the external world is also important.
DT interoperability will also be important in this DT Metaverse. In a city-oriented
DT Metaverse, DTIs of various vehicles could participate. We would want to have
them interoperate with DTIs of roads, traffic signals, and other aspects of the envi-
ronment. This is so that we could have accurate predictions of what would happen
in this particular DT Metaverse as different vehicle DTIs moved about.
We could use multiple participant immersive DT Metaverses. We might have
a factory machine that is having problems. Experts from all over the world could
congregate in the DT Metaverse of the factory to troubleshoot the problem and
recommend a physical fix. These experts could adjust parameters within the DT
Metaverse of the DTI machine that would then send those commands to the actual
machine itself. This is in essence merging the physical and digital worlds.
Human participants could have meta capabilities. So not only could they be
looking at the DTI of the machine, but they could also look within the machine
to see certain elements of it so that they had a better understanding of what was
transpiring.
Obviously, the machines would need to be instrumented to provide this data, but
that is fully within the realm of possibility. While dealing with a PT, there is no
other choice but to be synchronous. However, these avatars could step into one of
the sub-virtual spaces to speed up the process to see what was going to happen and
then transition back into the main DT Metaverse space in order to take action. This
looks very much like the original Digital Twin model in Figs. 3 and 4.

5.2 Artificial Intelligence, Simulation, Digital Twins,

and Metaverse

We need to consider the role that Artificial Intelligence (AI) will have in the evolution
to a DT Metaverse. It has been asserted that we are moving into an era of Intelligent
Digital Twins (IDT) that have the characteristics of being active, online, goal-seeking,
and anticipatory [37]. Grieves claim for DTs and IDTs is that AI will be an assistance
for humans and not a replacement.
It should be obvious that as we move into DTM platforms with major requirements
for replicative and predictive analysis and simulation and ensuing adverse prevention
that humans will need help. That help will take the form of AI.
In an above section, we discussed the difference between humans and nature.
While humans are task goal oriented while attempting to minimize physical
resources, nature tries all possible combinations via genetic mutations, letting the
environment select for fitness.
28 M. Grieves and E. Y. Hua

While the habitat of nature is the physical world, the habitat of AI is the digital
world or the DT Metaverses. Subject to computational limits and time constraints, AI
has the ability to explore the solution space, test it against the criteria for obtaining
goal success. AI can then select those solutions that meet or exceed the success
criteria.
In order to evolve to Phase 4 with Front Running Simulation, we are going to
need AI capability. FRS will require massive amounts of data coming from DTIs on
a constant basis. FRS will then require selecting the assumptions and parameters and
running simulations of future outcomes. The probabilities of those outcomes will be
calculated from the DTA and its causal and Bayesian probabilities.
However, it’s simply not enough to have probabilities of adverse events. We will
want to take action to prevent them from occurring. From the beginning days of DTs,
a key characteristic was cued availability [22]. Cued availability is a digital assistant
that assesses what is happening in context and offers information that replaces wasted
physical resources. Clearly an adverse event is an event that wastes resources.
Cued availability is a role for AI. Some adverse events will have obvious solutions.
It is obvious that the remediation for the prediction that there is a high probability of
a component failure in few weeks is to replace the component before the predicted
failure. Other predictions of adverse events will not have obvious remediations. An
example is a robot glitching momentarily that will result in an assembly bottleneck
later that day. AI has the ability to propose alternative preventative solutions for the
different adverse events it predicts.
This is clearly beyond human capability, computability, and mental bandwidth.
This will require the assistance of an AI agent with an enormous amount of computing
capability at its disposal. We may delegate some autonomous operations to AI
systems in FRS. However, we should always understand that AI should be an assistant
to humans and not a replacement.

6 Conclusion

As we have emphasized, we are in the twenty-first century. However, the movement

of work into the digital arena as described here is so dramatic that we may want to
refer to this as the 3rd millennium to call out how different we are from previous eras
that did not possess the computational and communication capability we now have
at our disposal. This century/millennium has seen the rise of technological capability
that did not exist or was immature prior to 2000. This includes technologies that we
are discussing here such as Digital Twins, photorealistic simulations, hybrid physics/
big data models, and AI.
While we have always had simulations and Replica Twins, the advances in
computing and communication technologies have enabled us to move work into
the digital world and use the information developed there to replace wasted physical
Defining, Exploring, and Simulating the Digital Twin Metaverses 29

resources in our goal-oriented tasks. We can use Digital Twins to replicate their phys-
ical counterparts so that we have simultaneous and instantaneous access to their data,
develop that data into information, and then predict their behavior via simulation.
The idea of a DT Metaverse will allow us to continue to evolve this capability
as we move to Phase 3 and 4 DT platforms. However, the DT Metaverse will need
the specific characteristics that we’ve outlined, such as fidelity to our universal laws.
Having our products employ Front Running Simulations will be our crystal ball into
the future, with the opportunity of using AI to prevent adverse events.
This may seem fantastical until we realize that it was only fifty years ago that
computer chip capacity was measured in thousands of bytes and that we recently
passed 100 billion bytes on a chip, with more exponential increases to come in the
next decade or so. With that in mind, the advances proposed in this chapter seem
very reachable and promise to create substantial value.

References

1. Lu H (2024) A framework for the credibility evaluation of equipment digital twins. In: Grieves
M, Hua E (eds) Digital twins, simulation, and metaverse: driving efficiency and effectiveness in
the physical world through simulation in the virtual worlds, Forthcoming. Springer, Heidelberg
2. Vickers J (2024) NASA’s Michoud assembly facility: developing a factory digital twin. In:
Grieves M, Hua E (eds) Digital twins, simulation, and metaverse: driving efficiency and effec-
tiveness in the physical world through simulation in the virtual worlds, Forthcoming. Springer,
Heidelberg
3. Malik AA (2024) Digital twins for robot systems in manufacturing. In: Grieves M, Hua E (eds)
Digital twins, simulation, and metaverse: driving efficiency and effectiveness in the physical
world through simulation in the virtual worlds, Forthcoming. Springer, Heidelberg
4. Boschert S (2024) On the importance of simulation and digital twin for industrial applications.
In: Grieves M, Hua E (eds) Digital twins, simulation, and metaverse: driving efficiency and
effectiveness in the physical world through simulation in the virtual worlds, Forthcoming.
Springer, Heidelberg
5. Darrington J (2024) DThe first real-time digital twin of a nuclear reactor. In: Grieves M, Hua
E (eds) Digital twins, simulation, and metaverse: driving efficiency and effectiveness in the
physical world through simulation in the virtual worlds, Forthcoming. Springer, Heidelberg
6. Dong Y (2024) Evolution of simulation and digital twin in healthcare: from discovery to design
and integration. In: Grieves M, Hua E (eds) Digital twins, simulation, and metaverse: driving
efficiency and effectiveness in the physical world through simulation in the virtual worlds,
Forthcoming. Springer, Heidelberg
7. Alexandridis K (2024) Distributed AI modeling and simulation for smart airport digital twin
applications. In: Grieves M, Hua E (eds) Digital twins, simulation, and metaverse: driving
efficiency and effectiveness in the physical world through simulation in the virtual worlds,
Forthcoming. Springer, Heidelberg
8. Moradian A (2024) Digital twins for sustainable semiconductor manufacturing. In: Grieves M,
Hua E (eds) Digital twins, simulation, and metaverse: driving efficiency and effectiveness in
the physical world through simulation in the virtual worlds. Editors, Forthcoming. Springer,
Heidelberg.
9. Mondoloni S (2024) The power and application of digital twins: a multidisciplinary perspective.
In: Grieves M, Hua E (eds) Digital twins, simulation, and metaverse: driving efficiency and
effectiveness in the physical world through simulation in the virtual worlds, Forthcoming.
Springer, Heidelberg
30 M. Grieves and E. Y. Hua

10. Singh S (2024) Defining, exploring, and simulating the digital twin metaverses. In: Grieves M,
Hua E (eds) Digital twins, simulation, and metaverse: driving efficiency and effectiveness in
the physical world through simulation in the virtual worlds, Forthcoming. Springer, Heidelberg
11. Hartmann S (1996) The world as a process: simulations in the natural and social sciences.
Modelling and simulation in the social sciences from the philosophy of science point of view.
Springer, Heidelberg, pp 77–100
12. Tzu S (2000) Art of War
13. Allison D (2019) Scotland’s role in the D-Day landings [cited 2023 August 23]. Available
from: https://www.bbc.com/news/uk-scotland-48266481
14. Barzun J (2000) From dawn to decadence: 500 years of cultural triumph and defeat, 1500 to
the present. HarperCollins
15. Schriber TJ (1974) Simulation using GPSS. Wiley, New York, xv, 533 p
16. Cass D, Shell K (1983) Do sunspots matter? J Polit Econ 91(2):193–227
17. Korb KB, Nicholson AE (2010) Bayesian artificial intelligence. CRC Press
18. Pearl J (2003) Statistics and causal inference: a review. TEST 12:281–345
19. Guo Y et al (2023) Digital twins for electro-physical, chemical, and photonic processes. CIRP
annals manufacturing technology, Forthcoming
20. Grieves M (2019) Virtually intelligent product systems: digital and physical twins. In: Flumer-
felt S et al (eds) Complex systems engineering: theory and practice. American Institute of
Aeronautics and Astronautics, pp 175–200
21. Grieves M (2024) DIKW as a general and digital twin action framework: data, information,
knowledge, and wisdom. Knowledge, Forthcoming
22. Grieves M (2006) Product lifecycle management: driving the next generation of lean thinking.
McGraw-Hill, New York
23. Popper K (2014) Conjectures and refutations: the growth of scientific knowledge. Routledge
24. Pillsbury J (2015) Modeling the world: ancient architectural models now on view [cited 2022
September 7, 2022]. Available from: https://www.metmuseum.org/blogs/now-at-the-met/2015/
modeling-the-world-ancient-architectural-models
25. New Link Trainer (1936) Aviation. McGraw-Hill Publishing Co, p 3
26. Furet J (1985) New concepts in control-room design: Worldwide, more attention is being given
to improving the man-machine interface. In: Nuclear power and electronics
27. Grieves M (2023) Digital twins: past, present, and future. In: Crespi N, Drobot AT, Minerva R
(eds) The digital twin. Springer International Publishing, Cham, pp 97–121
28. Holland S (2022)Michael Grieves, Father of Digital Twins, to headline at IOTSWC. In: iot
insider
29. Grieves M (2002) Completing the cycle: using PLM information in the sales and service
functions [Slides]. In: SME management forum, Troy, MI
30. Caruso P, Dumbacher D, Grieves M (2010) Product lifecycle management and the quest for
sustainable space explorations. In: AIAA SPACE 2010 conference & exposition, Anaheim,
CA
31. Grieves M (2011) Virtually perfect: driving innovative and lean products through product
lifecycle management. Space Coast Press, Cocoa Beach, FL
32. Grieves M (2014) Digital twin: manufacturing excellence through virtual factory replica-
tion (White Paper). Available from: https://www.3ds.com/fileadmin/PRODUCTS-SERVICES/
DELMIA/PDF/Whitepaper/DELMIA-APRISO-Digital-Twin-Whitepaper.pdf
33. Grieves M (2015) Can the digital twin transform manufacturing. Available from: https://www.
weforum.org/agenda/2015/10/can-the-digital-twin-transform-manufacturing/
34. Brown C, Singer T (2019) Defining and utilizing the digital twin. In: Digital enterprise society
35. Grieves M (2023) Digital twin certified: employing virtual testing of digital twins in
manufacturing to ensure quality products. Machines 11(808)
36. Andres Arcones D (2023) Evaluation of model bias identification approaches based on Bayesian
inference and applications to digital twins
37. Grieves M (2022) Intelligent digital twins and the development and management of complex
systems. Digital Twin 2(8)
Defining, Exploring, and Simulating the Digital Twin Metaverses 31

38. Oren T et al (2023) Simulation as experimentation. In: Oren T, Zeigler BP, Tolk A (eds)
Body of knowledge for modeling and simulation: a handbook by the society for modeling and
simulation. Springer Nature, pp 77–119
39. Grieves M (2006) Product lifecycle management : driving the next generation of lean thinking.
McGraw-Hill, New York, xiii, 319 p
40. Stephenson N (1991) Snow crash: a novel. Bantam Books
41. Bell MW (2008) Toward a definition of. J Virtual Worlds Res 1(1)
42. Joshua J (2017) Information bodies: computational anxiety in neal stephenson’s snow crash.
Interdisc Literary Stud 19(1):17–47
43. Rymaszewski M (2007) Second life: the official guide. John Wiley & Sons
44. Bohné T, Li K, Triantafyllidis K (2023) Exploring the industrial metaverse: a roadmap to the
future. World Economic Forum and Univesity of Cambridge
45. MIT Technology Review Insights (2023) The emergent industrial metaverse
46. Carroll S (2020) Something deeply hidden: Quantum worlds and the emergence of spacetime.
Penguin
47. Parker G, Van Alstyne M, Choudary SP (2016) Platform revolution: how networked markets
are transforming the economy and how to make them work for you, 1st edn. W.W. NORTON &
Company, New York, xiii, 336 pages
48. Richter S, Richter A (2023) What is novel about the Metaverse? Int J Inf Manage 73:102684
Contrasting Capabilities: A Comparative
Analysis of Simulation and Digital Twin
Technologies

Ana Wooley and Daniel F. Silva

Abstract In the realm of advanced manufacturing, confusion often surrounds the

distinction between Digital Twin (DT) and simulation due to their complementary
nature. While they appear similar, they are fundamentally different concepts. This
chapter considers capability-based frameworks to characterize and classify DT and
simulation. Specifically, it uses the 4R framework for DT, from the literature, and
proposes an analogous one (4S framework) for simulation. By comparing the defi-
nitions and levels of capability of DT and simulation using these frameworks, this
chapter identifies similarities and differences in capability between these two tech-
nologies. Ultimately, this chapter aims to clarify these differences and provide a
comprehensive comparison between the two, to prevent future ambiguity on this
topic. The conclusion of this study is that differences in capabilities between the two
technologies can be observed in data flows, real-time aspects, predictive timelines,
optimization approaches, and level of autonomy.

Keywords Simulation · Digital Twin · Capability

1 Introduction

In the advanced manufacturing literature, there is often confusion surrounding the

distinction between Digital Twins (DTs) and simulations. This confusion arises from
their complementary nature, leading to the perception that they are similar or inter-
changeable concepts. However, it is important to recognize that DTs and simulations
are fundamentally different. To address this issue, this chapter aims to clarify the
differences between the two by exploring their individual capabilities and presenting
a comprehensive comparison between simulation and DTs. By providing a clear

A. Wooley (B)
The University of Alabama in Huntsville, Huntsville, USA
e-mail: [email protected]
D. F. Silva
Auburn University, Auburn, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 33

M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_2
34 A. Wooley and D. F. Silva

comparison between DTs and simulations, this chapter emphasizes the synergistic
relationship between them while emphasizing their distinct characteristics and capa-
bilities, enabling informed decision-making and effective implementation of these
technologies in various industries.
The chapter begins by defining the concept of a DT, introducing the 4R framework
[1, 2] to characterize and classify DTs based on levels of capability. Next, it presents
the definition of simulation and proposes an analogous framework to characterize
simulation based on its levels of capability. This chapter highlights that a simulation
can be based on reality, or it can be completely artificial, with the primary objective
centered on examining how an object or system evolves and produces outcomes over
time, using a predefined set of inputs. On the other hand, a DT fully replicates, in a
virtual environment, a physical element that exists or will exist in the physical world,
with bidirectional data flow connecting the physical and the virtual environments.
The main result presented here is a discussion of the similarities and differences
between DT and simulation, based on the capability-based classification frameworks
discussed earlier.
In summary, the main conclusion is that simulation and DT share a common foun-
dation of modeling and analysis but diverge in their objectives and advanced features.
DTs, unlike simulations, offer real-time adjustments and autonomy, enhancing
decision-making and optimizing physical systems.

2 Digital Twin Capabilities

The concept of DT has gained significant attention from both academia and industry
in recent years [3, 4]. This surge in interest in DT is primarily driven by Industry 4.0
and advancements in technologies, such as the Internet of Things (IoT), Artificial
Intelligence (AI), wireless sensor networks, Machine Learning (ML), and big data
[5]. These technologies offer opportunities to seamlessly merge physical and virtual
environments. The growing interest in DT has also led to an increase of publications
about it [6, 7]. However, even academic publications about DT do not always agree
on the what the definition, characteristics, and capabilities of DT are, or should be,
resulting in a dilution of the DT concept [8–10], and causing confusion regarding its
practical applications and advantages [11].
The concept of DT was originally introduced by Michael Grieves in collaboration
with John Vickers from the National Aeronautics and Space Administration (NASA)
[12]. Grieves’s definition of a DT consists of three key elements: a physical space,
a virtual space, and the data flow that connects these spaces bidirectionally. Since
then, researchers have been trying to establish a unified and universally accepted
definition that effectively captures the essence of a DT [6, 10]. Some researchers,
as exemplified by [13–16], focus on DTs applied to individual products, where a
DT represents a single machine or process. Conversely, others, including [17–21],
consider DTs in the context of processes, encompassing either a segment or the
entirety of the production environment.
Contrasting Capabilities: A Comparative Analysis of Simulation … 35

In the community’s journey to determine the conclusive definition and capabilities

of a DT, some researchers, such as [22, 23], have presented concepts, definitions, and
architectures related to DT, while others, such as [24, 25], have delved into emerging
technologies relevant to industrial DT applications. Alternatively, [26] proposes that
the DT consists of different types (DT Prototype, DT Aggregate, and DT Instance)
for the entire product lifecycle.
Although certain authors define DT as a model that integrates both the physical
system and its virtual counterpart with a continuous and bidirectional data exchange
between them [21], numerous other definitions persist, characterizing DTs solely
as accurate representations of the physical environment, disregarding real-time data
connections [27]. To achieve a consensus on a DT definition, the specifications of
fundamental requirements and capabilities for a DT are necessary. Furthermore, these
requirements may change over time because they are linked to advancements in the
technologies, such as ML and big data [27].
Several studies have focused on establishing a framework for DTs and explaining
the interactions among various DT components [1, 2, 28–31]. Each of these works has
attempted to provide a definitive methodology for characterizing DTs. While each
of the proposed approaches has advantages, we have chosen to focus on a single
taxonomy of DT throughout this chapter to provide a direct comparison between
DTs and simulation models.
Specifically, we will use a robust, versatile, and systematic framework for charac-
terizing DTs based on their capability and maturity levels called the 4R framework [1,
2]. This framework comprises four distinct levels of DT: Representation (R1), Repli-
cation (R2), Reality (R3), and Relational (R4). Any DT can be classified into one
of the 4R levels based on its maturity and capabilities, which increase in each level.
Additionally, each level within this framework outlines the developmental process
of a DT for a given system.
Representation (R1): In the 4R framework, the first level marks the first stage of
building a DT. It involves understanding the operations of the physical system and
establishing a mechanism for gathering and storing data derived from the physical
surroundings. At this level, data is connected and employed to construct a represen-
tation of the physical environment. This data serves the purposes of visualization,
validation, and the control of the virtual environment.
Replication (R2): In this level, the emphasis shifts toward replicating the system
within a virtual setting using the architecture created during the Representation phase.
At this level, the DT has the ability to replicate identical outcomes when provided
with identical inputs as those of the physical system.
Reality (R3): The third level is where the DT is deployed to explore what-if
scenarios, with the goal of leveraging the insights gained from virtual models to
enhance the performance of the physical system. At this level, the DT model possesses
the capability to generate results based on a set of inputs, operating independently
of the physical system.
Relational (R4): The last level describes a DT with the capabilities of integrating
decision-making technologies by using AI or ML. At this level, the physical and
36 A. Wooley and D. F. Silva

Fig. 1 Summary of 4R framework. Figure adapted from [1, 2]

virtual systems are connected and the data between them is bi-directional. Addition-
ally, the DT has a level of autonomy that is not present in the preceding capability
levels, enabling it to autonomously adapt, identify optimal strategies, and fine-tune
itself.
A concise overview of the DT capabilities is depicted in Fig. 1.

3 Simulation Capabilities

Simulation is a modeling technique that replicates the operations of a process or

system over time, serving the purposes of describing and analyzing the system
behavior and conducting what-if scenarios [32]. Simulation offers a host of advan-
tages, providing methodologies and tools for intricately modeling system behaviors,
facilitating predictions of future outcomes, and thereby supporting decision-making
procedures [33]. It enables the testing of designs without committing resources to
physical acquisition, as well as the ability to manipulate and accelerate time, which
is relevant in system investigations [34].
Simulation also provides understanding of why things happen the way they do in
the system and explores possibilities without the expense and disruption of experi-
menting with the real system [35]. In essence, a simulation model serves as a precise
representation of a system or subsystem that fulfills specific criteria, frequently used
to verify and analyze potential improvements prior to their actual implementation in
a real-world setting [36].
The utilization of simulation technology is a well-established practice in the field
of engineering [37], offering a wide array of simulation paradigms to choose from
[38]. The selection of the most appropriate paradigm to apply when simulating a
system depends on several factors, including the characteristics of the system, specific
requirements that must be met by the simulation, and the desired outcomes of the
Contrasting Capabilities: A Comparative Analysis of Simulation … 37

analysis. The three most prevalent simulation paradigms are Monte Carlo, Agent-
based, and Discrete Event Simulation (DES) [39].
In simple terms, a Monte Carlo simulation is employed to assess uncertainties and
potential risks by means of probability distributions [40]. Monte Carlo simulation is
often used to generate random variates that will be utilized in more complex analysis,
such as Monte Carlo Markov Chain models.
In contrast, an Agent-based simulation models individual entities (agents) and
their interactions with each other and the environment, facilitating the modeling of
intricate systems with numerous interacting components [41]. In Agent-based simu-
lations, agents act in a decentralized manner, following a set of prescribed (possibly
randomized) rules that govern their interactions with other entities.
DES, on the other hand, permits the modeler to generate or observe specific
events over time and it proves versatile for studying various system types and an
extensive range of outcomes [36]. DES stands as the most used approach within the
manufacturing domain [42].
Extensive literature exists defining simulation and offering detailed instructions
for constructing simulation models. Many authors agree [35, 43–46] that simula-
tion models provide the capability of conducting experiments and analyzing results.
However, none of them provide a structured framework for characterizing and clas-
sifying simulations that would allow us to distinguish between the different levels
of capability or that can be directly compared to other technologies (by comparing
capabilities), such as DT.
While numerous frameworks and guidelines exist in the literature for constructing
simulations, to the best of our knowledge, there is no framework that precisely delin-
eates the different levels of capability within a simulation model in a manner that is
comparable to a framework for DT, such as the 4R framework. To be clear, the capa-
bilities of simulations are thoroughly documented in existing literature; therefore, the
objective of developing such a framework is to consolidate these well-established
definitions and integrate them into a comprehensive structure that can serve as a
roadmap for understanding simulation models based on their capabilities, which is
analogous to the existing 4R framework for DTs.
An analogous framework for evaluating simulation capabilities should provide
a similar assessment of the level of simulation maturity in an organization. This
would enable organizations to gauge their current simulation capabilities, pinpoint
areas requiring improvements, and develop a plan for future simulation investments.
Furthermore, an analogous framework would facilitate comparisons between simu-
lation and DT capabilities, allowing organizations to better align their simulation and
DT strategies.
In this context, we present a framework for simulation, which we called the 4S
framework, that was initially introduced in our prior work [47]. It is designed to be
analogous to the 4R framework for DTs, allowing for the classification of applications
of both DT and simulation and facilitating comparisons between them. It is important
to emphasize that there isn’t a one-to-one correspondence between the levels in the
4S and 4R frameworks.
38 A. Wooley and D. F. Silva

Fig. 2 Systematic search for textbooks and articles. Figure adapted from [47]

To inform the 4S framework’s development, definitions of simulation were gath-

ered from research literature dedicated to simulation and its characteristics. Specifi-
cally, a two-pronged literature search was conducted: one for textbooks and another
for scholarly articles. The selection process for the textbooks and articles used in this
research is outlined in Fig. 2.
To gather pertinent textbooks, three prominent academic publishers were chosen:
CRC Press, John Wiley & Sons, and McGraw Hill. These publishers are renowned
for their expertise in producing textbooks and reference materials and enjoy a
strong reputation for publishing content across a broad spectrum of academic disci-
plines. They hold a respected position within the academic community, particularly
known for their authoritative publications in areas like engineering, technology, and
computer science, which are directly relevant to the study of simulation. By focusing
on these publishers, the likelihood of identifying relevant books that provide thor-
ough insights and analyses regarding simulation and its attributes is significantly
enhanced.
The first step involved conducting a search to identify textbooks featuring the
term “simulation” in their titles. To maintain relevance to manufacturing and service
system applications, this search was confined to books that listed “industrial engi-
neering” as one of its topic areas. From these databases, CRC Press (25 books),
John Wiley & Sons (8 books), and McGraw Hill (2 books), a total of 35 books were
retrieved. Subsequently, a thorough and rigorous evaluation process was executed
to select the most suitable books for this study. This process entailed sorting the
results based on relevance and analyzing the titles of each book, with a focus on
those directly providing simulation concepts.
The criteria for inclusion and exclusion were carefully designed to consider
only books encompassing the general theory of simulation, such as handbooks,
which incorporate definitions and characteristics of simulation. Any books explicitly
mentioning a specific simulation type or application field in their titles were excluded
to prevent any bias toward a particular definition of simulation associated with a
specific simulation paradigm or application area. These criteria were established to
ensure that the focus remains solely on the core theory of simulation and its funda-
mental principles, avoiding any influence from specific simulation types like DES
Contrasting Capabilities: A Comparative Analysis of Simulation … 39

or Agent-based simulation. Following the initial assessment of book titles, 19 books

were found to not meet the inclusion criteria and were excluded. As a result, six text-
books [35, 36, 48–51] were identified as pertinent sources for gathering definitions
of simulation.
To obtain relevant literature for the article search, the database of choice was the
Winter Simulation Conference (WSC) proceedings. This decision was motivated by
the prominence of WSC as a conference that consistently features a track dedicated
to introductory tutorials, which cover the subject of simulation and its fundamental
characteristics. These tutorials are typically authored by leading experts in the field,
offering comprehensive insights into the concept and definition of simulation. Such
an emphasis on definitions and concepts is not commonly encountered in journal
articles, which often assume a pre-existing familiarity with the field and focus only
on novel, recent developments.
A filter was applied to identify articles related to a general introduction to simu-
lation, published within the last 25 years in the introductory tutorial track, resulting
in the retrieval of 20 publications. It’s worth noting that some of these articles were
written by the same author in consecutive years. In such cases, the most recent
versions were selected, as they often represented refined iterations of what had been
presented the previous year. Consequently, eight pertinent papers were obtained
[43, 45, 52–57].
The next step involved the aggregation and examination of definitions and
attributes of simulation presented in the selected books and articles using text
analytics. The definitions were transcribed verbatim into a spreadsheet. Next, each
phrase within every definition was isolated into individual cells within the spreadsheet
and characterized by identifying the action verb within each phrase.
By deconstructing the definitions into their component parts and clustering them
based on common themes, six main characteristics of simulation were identified
and labeled as C1 through C6. For instance, many phrases featured the action verb
“understand,” implying a focus on understanding a specific system or process. By
grouping together phrases utilizing similar action verbs, a characteristic denoted
as “Understanding how the system works” was established and designated as C1.
Similarly, action verbs like “build” and “create” appeared frequently in numerous
phrases, indicating an emphasis on simulation’s capability to construct a model of
a system. These phrases were categorized as C2, labeled as “Represent the physical
system in a digital model.” A comparable approach was employed to arrive at C3
through C6. A summary of these characteristics is provided in Table 1.
The subsequent step involved revisiting the books and articles to identify which of
the characteristics (C1–C6) were explicitly mentioned in their definitions of simula-
tion. A summary of this classification process is presented in Table 2, which illustrates
that while certain characteristics are implied, some authors do not expressly articu-
late them in their definitions, as is the case with C1, C3, and C4. On the other hand,
all sources explicitly state that simulation is a digital representation of a system
(C2), it is employed to evaluate and predict the behavior of physical systems under
varying conditions (C5), and it serves as a source of insights for decision-making
(C6). Consequently, after an examination of the individual definitions pertaining to
40 A. Wooley and D. F. Silva

Table 1 Summary of the characteristics of simulation

Characteristics
(C1) Understand how the system works
(C2) Represent the physical system in a digital model
(C3) Examine and analyze the system and its components
(C4) Draw inferences about the system
(C5) Evaluate and predict how the physical system will behave under different conditions
(C6) Provide insight into the system’s optimal operation to support decision-making
Table adapted from [47]

each characteristic (C1–C6), it was found that certain categories exhibited significant
conceptual overlap. As a result, C1 and C2 were consolidated into a single category,
and C3 and C4 were likewise combined.
The categories “Understand how the system works” (C1) and “Represent the
physical system in a digital model” (C2) were combined into a single category
because many authors combine understanding and representation of a system into one
step [35, 36, 43, 49–51, 53, 54]. C1 emphasizes the need for understanding a system’s
behavior to create an accurate simulation model. This encompasses the identification
of various system components, their interactions, and their interrelationships [58].
On the other hand, C2 involves the development of a digital representation of a
physical system. This process involves constructing a simulation model that faithfully

Table 2 Simulation definition sources classification

Ref C1 C2 C3 C4 C5 C6
Books [36] ✕ ✕ ✕ ✕ ✕ ✕
[35] ✕ ✕ ✕ ✕ ✕ ✕
[48] ✕ ✕ ✕ ✕ ✕
[49] ✕ ✕ ✕ ✕ ✕
[50] ✕ ✕ ✕ ✕ ✕
[51] ✕ ✕ ✕ ✕ ✕
Articles [52] ✕ ✕ ✕ ✕
[53] ✕ ✕ ✕ ✕ ✕ ✕
[54] ✕ ✕ ✕ ✕ ✕
[43] ✕ ✕ ✕ ✕ ✕
[45] ✕ ✕ ✕ ✕ ✕
[55] ✕ ✕ ✕ ✕
[56] ✕ ✕ ✕ ✕
[57] ✕ ✕ ✕ ✕
S1 S2 S3 S4
Table ad apted from [47]
Contrasting Capabilities: A Comparative Analysis of Simulation … 41

mirrors the characteristics and behavior of the physical system. To accomplish this,
the simulation modeler must have an in-depth understanding of the behavior of the
physical system, its components, and their interactions [33].
The categories “Examine and analyze the system and its components” (C3) and
“Draw inferences about the system” (C4) were combined into a single category
because both involve the process of analyzing a system to draw conclusions about
its behavior. C3 entails the examination and analysis of various system components
within the simulation model to understand how they interact with one another. This
analysis encompasses the identification of relationships among different components
and their impact on the overall system behavior. Similarly, C4 is centered around
drawing conclusions about the system’s behavior based on the outcomes generated
by the simulation model. Upon the completion of the simulation, the modeler can
analyze the results and draw conclusions regarding the system’s behavior [59].
After combining similar characteristics, the resulting framework comprises four
categories. To make it more comparable it with the 4R framework of DT, these
categories are defined as the 4S framework of simulation: S1 (combining C1 and
C2), S2 (combining C3 and C4), S3 (C5), and S4 (C6). To appropriately label the
four categories in the 4S framework of simulation, the most common and relevant
terms associated with each category were visually represented in a word cloud. This
approach helped to simplify and condense the keywords, making it easier to identify
the most important and frequently used terms. The word clouds for each category
are shown in Fig. 3.
Following a thorough analysis of the word clouds, names were assigned to each
of the four categories in the 4S framework of simulation. To encapsulate the essence
of each category as a capability, it was decided to use action verbs when selecting
category labels. In the word cloud of the first category (S1), the most frequent terms

Fig. 3 Word Clouds for labeling the 4S levels. Figure adapted from [47]
42 A. Wooley and D. F. Silva

included “model,” “behavior,” “understand,” and “real.“ “Model” and “behavior”

strongly imply a focus on simulating the behavior of a system. Similarly, “under-
stand” and “real” indicate an emphasis on comprehending the workings of the actual
system. These terms are closely associated with conceptual and computer models,
suggesting that “Modeling” is a fitting title for this category.
In the word cloud for the second category (S2), the most prevalent terms were
“analyze” and “behavior”. These terms suggest that this category centers on the
analysis and assessment of system behavior and performance, making “Analyzing”
an appropriate name for this category.
In the word cloud for the third category (S3), the most frequently recurring terms
were “predict,” “hypothesis,” “what-if,” “test,” and “evaluate”. These terms imply that
this category concentrates on using simulation to forecast how a system will behave
under different conditions and to test and evaluate what-if scenarios. Therefore,
“Predicting” serves as a fitting designation for this category.
Lastly, in the word cloud of the fourth category (S4), the most frequent terms
included “decision,” “making,” “provide,” “insights,” and “recommendations.” These
terms suggest that this category revolves around employing simulation to provide
valuable insights and recommendations that can inform decision-making. The term
“prescriptive” is commonly employed in the realm of computing and analytics to
refer to a type of analytics that involves recommending specific actions to achieve
a desired outcome [60]. Therefore, “Prescribing” is an appropriate label for this
category.
Overall, as the level of a simulation increases from S1 to S4, its level of analytical
utility and its capabilities both increase [47]. A summary of these levels of capability
is shown in Fig. 4.
Modeling (S1): The first level involves creating an abstracted representation of a
real-world system through conceptual modeling and computer modeling. Conceptual
modeling helps in understanding the system, mapping its processes, and collecting

Fig. 4 Summary of the simulation capabilities (4S Framework). Figure adapted from [47]
Contrasting Capabilities: A Comparative Analysis of Simulation … 43

requirements. In the computer modeling stage, the system components are repre-
sented in a virtual format and the simulation model is created. Proper data collection,
parameter selection, and understanding are crucial in this stage.
Analyzing (S2): In this level, the data gathered during modeling is incorporated
into the simulation model. Input analysis ensures that the data used approximates real-
world behavior. Verification and validation are performed to ensure model accuracy.
In this phase, analysis is conducted to draw inferences about the system’s behavior.
Predicting (S3): This level is where the simulation has the capability to predict
system behavior under various conditions. It allows for the investigation of alter-
native scenarios, performance measurement, and comparison. Developing accurate
scenarios is essential.
Prescribing (S4): This level involves using the simulation model for decision
support and optimization. Decision-makers can identify optimal courses of action,
and simulation optimization or machine learning methods can be applied to enhance
decision-making and system performance.

4 What Is the Difference Between Simulation and Digital

Twin?

Although traditional simulation and DT models share similar capabilities in repli-

cating physical systems within a virtual environment, they are not the same and
should not be referred to as the same technology [61]. The practice of combining
simulation and DT can lead to confusion when simulation models are mislabeled as
DT and vice versa [62]. While traditional simulation serves a valuable role in various
applications, it is not accurate to label it as a DT.
Establishing a clear distinction between the capabilities of a DT and a simulation
is essential to maintain consistency in DT implementations. DTs and simulations
represent distinct technologies, each offering unique advantages that can yield valu-
able insights into various problems. Therefore, it is important to accurately categorize
them and clarify their differences to prevent misconceptions.
In prior work, we conducted a systematic literature review [63] with the primary
objective of determining whether the current body of literature, that claims to combine
simulation and DT, is truly applying DT or whether it is using simulation in place of
DT. Figure 5 shows the findings from this study. It indicates that one third of the papers
(33%) are classified as R0 because they did not achieve any of the DT capabilities
according to the 4R framework. These works simply presented simulations and called
them DT. This demonstrates a persistent gap between the conceptual understanding
of DT and its practical applications as many individuals continue to build models
and label them as DTs even though they are essentially simulation models. Even
among those who build DTs correctly, they often harness only a fraction of their
capabilities, failing to realize their full potential, with no publications among those
surveyed reaching the R4 level.
44 A. Wooley and D. F. Silva

60
2
50 9
40 2
13
30 19
20
30
13 9
10
6 9 8
0
R0 (33%) R1 (8%) R2 (45%) R3 (14%) R4 (0%)

S1 S2 S3 S4

Fig. 5 Analysis of the systematic literature review. Figure adapted from [63]

With the introduction of the 4S framework for simulation and the existing 4R
framework for DTs, we can now proceed with a direct comparison of both technolo-
gies. In the following discussion, we will explore the similarities and differences
between simulation and DTs.
. Digital Representation: Both simulation and DTs involve creating a virtual repre-
sentation of a physical system by gathering and understanding data from the
physical world. The key difference is that a simulation can vary from being very
realistic to highly abstracted, whereas DT fully describes an existing physical
element or one that will exist in the physical world [26].
. Real-Time: Simulations may rely on data-driven approaches, but they usually
lack real-time data integration. Even at the most advanced level of simulation
capability (S4), real-time data flows are typically absent. Simulation typically
relies on historical data and performing input analysis of the data to use proba-
bility distributions or rules to represent that system over time, whereas the DT is
continuously updated to mirror the real-world counterpart in real time. The real-
time capability of a DT happens during the R2 and R3 phase; however, the key
parameters and data pipeline of the DT happens during the first level (R1), which
requires the establishment of real-time data pipeline integration from sensors, IoT
devices, or other sources.
. Data Flows: Simulations usually involve one-way data flow, which is established
at the first capability level (S1). DTs, on the other hand, commonly feature bidi-
rectional data flow, with the flow from the physical to the virtual domain set up
during the representation phase (R1), and the virtual to physical data flow estab-
lished in the final capability level (R4). Bidirectional data flows are not a typical
component of simulations, even at higher capability levels (S2, S3, and S4).
. Analysis and Visualization: Both simulation and DTs can be used for analysis and
visualization, which involves using the data from the physical system to analyze
the system’s behavior and replicate it within a virtual environment. Both can
replicate identical outcomes when provided with the same inputs as those of the
Contrasting Capabilities: A Comparative Analysis of Simulation … 45

physical system. The distinction lies in their respective approaches. Simulations

often provide a description of the system based on historical data and require
extensive pre-analysis to explore how an object or system will change over time
under various input conditions. In contrast, DTs place a stronger emphasis on real-
time analysis and monitoring, focusing on the current state and performance of the
physical system. In other words, DTs can display real-time events as they occur,
providing insights into the present situation, past events, and future occurrences.
Simulations, on the other hand, can only represent what has happened or what
will happen in the future, lacking the ability to capture real-time.
. Prediction Timeframe: Both simulation and DTs possess the capability to predict
the behavior of a system under various scenarios, allowing for the exploration of
“what-if” scenarios. They can generate results based on a specific set of inputs,
operating independently from the physical system, and producing outcomes and
changes in state that mirror what the physical system would produce within the
same timeframes. However, simulations typically place a greater emphasis on
medium and long-term-oriented analysis, focusing on how a product evolves as it
interacts with external forces and how these forces impact the system over time,
which happens during the predicting level of capability of simulation (S3). In
contrast, DTs do not exclusively concentrate on predicting the future; they also
deliver real-time or near-real-time data concerning the current state and behavior
of the physical system, corresponding to the third level of capability of DTs
(R3). While simulation can give insight into the distant future and prescribe solu-
tions based on the system’s historical state, DTs can adapt to new situations and
reveal what is currently happening in the system. If a sudden change in conditions
happens, such as the disruption to global supply chains as a result of the Covid-19
pandemic, a simulation is not typically going to reflect it until a long time after the
fact, as it is usually fed historical data, which would not immediately register those
changes. On the other hand, a DT, which is focused on reproducing the physical
system in real-time, would quickly incorporate the effects of such a disruption
and be capable to extrapolate those impacts into the future.
. Autonomy: The predictive (S3) and prescriptive (S4) capabilities of simulation
are closely linked to the reality level of DTs (R3). Both simulation and DTs
are utilized to support decision-making and optimize processes. However, the
key distinction lies in the fact that simulations often depend on human operators
to make decisions and implement them based on the results generated by the
simulation. In contrast, DTs operate autonomously, reducing the need for direct
human intervention. In fact, at the R4 level, when data flows from the virtual world
to the physical world have been established, DTs can perform prescriptive tasks
and send commands directly to the physical system to implement those decisions.
In the context of DTs, they can provide prescriptions at the R3 level, and when
progressing to R4, automation is introduced.
. Optimization: Because DTs appeared at the same time as an explosion of research
and development of both ML and AI, DT architecture has been developed to
allow (at the R4 level) ML and AI to be directly integrated into DTs. This feature
is absent in most commercial simulation software. By incorporating advanced
46 A. Wooley and D. F. Silva

decision-making technologies such as AI and ML, jointly with bi-directional

data flows, DTs become capable of adapting, identifying optimal strategies, and
autonomously improving. By contrast, the prescriptive capabilities of simulation
(at the S4 level) are often done through calling external optimization software or
using iterative procedures between simulation and external tools. At best, most
commercial simulation software only provides the ability to fine-tune parameters
through enumerative approaches.
In summary, both simulation and DT start with modeling or representation and
progress through stages of analysis and replication. However, the key differences
lie in their data flows, the approach to real-time data adjustments, integration of
advanced technologies such as AI and ML, and autonomy.

5 Conclusion

In conclusion, this chapter addressed the prevalent confusion in the literature

regarding the distinctions between DTs and simulations, often misinterpreted as inter-
changeable concepts due to their complementary nature. It highlighted the critical
need to differentiate between DTs and simulations, emphasizing their fundamental
differences in capabilities. The chapter started by introducing an existing framework
for characterizing DTs based on their capability and maturity levels, called the 4R
framework. Next, the chapter presented the creation of an analogous framework to
characterize simulation based on its levels of capability, the 4S framework, allowing
for the classification of applications of both DT and simulation and facilitating direct
comparisons between their capabilities.
This chapter delineated the similarities and differences in capabilities between
simulations and DTs. The choice between simulations and DTs is not straightforward.
Nonetheless, DTs serve a broader array of purposes beyond running scenarios. While
both simulations and DTs employ digital models to replicate products and processes
within a virtual environment, several crucial distinctions set them apart.
The most significant difference between DT and simulation lies in the real-time
data integration and bidirectional data exchange in DTs, connecting the virtual envi-
ronment with sensors for continuous data collection. In contrast, simulations rely
on historical data and typically extensive pre-analysis, lacking real-time integration.
This enhanced connectivity in DTs enhances the precision of predictive analytical
models, offering a deeper understanding for product, policy, and procedure manage-
ment and monitoring. Furthermore, DTs emphasize bidirectional data flows, a feature
uncommon in simulations, and exhibit greater autonomy and advanced decision-
making technologies such as AI and ML, reducing the need for human interven-
tion. DTs can even autonomously command physical systems and boast superior
optimization capabilities compared to most commercial simulation software.
Contrasting Capabilities: A Comparative Analysis of Simulation … 47

References

1. Hyre A, Harris G, Osho J, Pantelidakis M, Mykoniatis K, Liu J (2022) Digital twins: Repre-
sentation, Replication, Reality, and Relational (4Rs). Manuf Lett 31:20–23. https://doi.org/10.
1016/j.mfglet.2021.12.004
2. Osho J, Hyre A, Pantelidakis M, Ledford A, Harris G, Liu J et al (2022) Four Rs frame-
work for the development of a digital twin: the implementation of representation with a FDM
manufacturing machine. J Manuf Syst 63:370–380. https://doi.org/10.1016/j.jmsy.2022.04.014
3. Liu M, Fang S, Dong H, Xu C (2021) Review of digital twin about concepts, technologies, and
industrial applications. J Manuf Syst 58:346–361. https://doi.org/10.1016/j.jmsy.2020.06.017
4. Ludwig S, Karrenbauer M, Fellan A, Schotten HD, Buhr H, Seetaraman S et al (2018) A 5G
architecture for the factory of the future 2018
5. Syafrudin M, Alfian G, Fitriyani NL, Rhee J (2018) Performance analysis of IoT-based sensor,
big data processing, and machine learning model for real-time monitoring system in automotive
manufacturing. Sensors (Switzerland) 18. https://doi.org/10.3390/s18092946
6. Tao F, Zhang H, Liu A, Nee AYC (2019) Digital twin in industry: state-of-the-art. IEEE Trans
Industr Inform 15:2405–2415. https://doi.org/10.1109/TII.2018.2873186
7. Melesse TY, Di Pasquale V, Riemma S (2021) Digital Twin models in industrial operations:
state-of-the-art and future research directions. IET Collaborative Intell Manuf 3:37–47
8. VanDerHorn E, Mahadevan S (2021) Digital Twin: generalization, characterization and
implementation. Decis Support Syst 145:113524. https://doi.org/10.1016/j.dss.2021.113524
9. Walton RB, Ciarallo FW (2022) A framework to manage systems based on digital twin. In: IIE
annual conference. Proceedings, Institute of Industrial and Systems Engineers (IISE), pp 1–6
10. Jones D, Snider C, Nassehi A, Yon J, Hicks B (2020) Characterising the digital twin: a system-
atic literature review. CIRP J Manuf Sci Technol 29:36–52. https://doi.org/10.1016/j.cirpj.2020.
02.002
11. Boschert S, Heinrich C, Rosen R (2018) Next generation digital twin. In: Proceedings of TMCE,
Las Palmas de Gran Canaria, Spain, pp 7–11
12. Grieves M, Vickers J (2017) Digital Twin: mitigating unpredictable, undesirable emergent
behavior in complex systems. In: Kahlen F-J, Flumerfelt S, Alves A (eds) Transdisci-
plinary perspectives on complex systems: new findings and approaches. Springer International
Publishing, Cham, pp 85–113. https://doi.org/10.1007/978-3-319-38756-7_4
13. Abramovici M, Göbel JC, Savarino P (2016) Virtual twins as integrative components of smart
products. In: Harik R, Rivest L, Bernard A, Eynard B, Bouras A (eds) Product lifecycle manage-
ment for digital transformation of industries. Springer International Publishing, Cham, pp
217–226
14. Schroeder GN, Steinmetz C, Pereira CE, Espindola DB (2016) Digital Twin data modeling with
automation ML and a communication methodology for data exchange. IFAC-PapersOnLine
49:12–17. https://doi.org/10.1016/j.ifacol.2016.11.115
15. Shao G, Kibira D (2018) Digital manufacturing: requirements and challenges for implementing
digital surrogates. In: 2018 Winter Simulation Conference (WSC), IEEE, pp 1226–1237
16. Tao F, Zhang M (2017) Digital Twin shop-floor: a new shop-floor paradigm towards smart
manufacturing. IEEE Access 5:20418–20427. https://doi.org/10.1109/ACCESS.2017.2756069
17. Gabor T, Belzner L, Kiermeier M, Beck MT, Neitz A (2016) A simulation-based architecture
for smart cyber-physical systems. IEEE Int Conf Autonomic Comput (ICAC) 2016:374–379
18. Rosen R, von Wichert G, Lo G, Bettenhausen KD (2015) About the importance of autonomy
and digital twins for the future of manufacturing. IFAC-PapersOnLine 48:567–572. https://doi.
org/10.1016/j.ifacol.2015.06.141
19. Coronado PDU, Lynn R, Louhichi W, Parto M, Wescoat E, Kurfess T (2018) Part data integra-
tion in the shop floor Digital Twin: mobile and cloud technologies to enable a manufacturing
execution system. J Manuf Syst 48:25–33
20. Negri E, Fumagalli L, Macchi M (2017) A review of the roles of digital twin in CPS-based
production systems. Proc Manuf 11:939–948
48 A. Wooley and D. F. Silva

21. Kritzinger W, Karner M, Traar G, Henjes J, Sihn W (2018) Digital Twin in manufacturing:
a categorical literature review and classification. IFAC-PapersOnLine 51:1016–1022. https://
doi.org/10.1016/j.ifacol.2018.08.474
22. Singh M, Fuenmayor E, Hinchy EP, Qiao Y, Murray N, Devine D (2021) Digital twin: origin
to future. Appl Syst Innov 4. https://doi.org/10.3390/asi4020036
23. Juarez MG, Botti VJ, Giret AS (2021) Digital twins: review and challenges. J Comput Inf Sci
Eng 21. https://doi.org/10.1115/1.4050244
24. Lim KYH, Zheng P, Chen CH (2020) A state-of-the-art survey of Digital Twin: techniques,
engineering product lifecycle management and business innovation perspectives. J Intell Manuf
31:1313–1337. https://doi.org/10.1007/s10845-019-01512-w
25. Rathore MM, Shah SA, Shukla D, Bentafat E, Bakiras S (2021) The role of AI, machine
learning, and big data in digital twinning: a systematic literature review, challenges, and
opportunities. IEEE Access 9:32030–32052. https://doi.org/10.1109/ACCESS.2021.3060863
26. Grieves MW (2023) Digital Twins: past, present, and future. The Digital Twin. Springer,
Heidelberg, pp 97–121
27. Barricelli BR, Casiraghi E, Fogli D (2019) A survey on Digital Twin: definitions, characteristics,
applications, and design implications. IEEE Access 7:167653–167671. https://doi.org/10.1109/
ACCESS.2019.2953499
28. Fan Y, Yang J, Chen J, Hu P, Wang X, Xu J et al (2021) A digital-twin visualized architecture
for flexible manufacturing system. J Manuf Syst 60:176–201. https://doi.org/10.1016/j.jmsy.
2021.05.010
29. Wu C, Zhou Y, Pereia Pessôa MV, Peng Q, Tan R (2021) Conceptual digital twin modeling
based on an integrated five-dimensional framework and TRIZ function model. J Manuf Syst
58:79–93. https://doi.org/10.1016/j.jmsy.2020.07.006
30. Zhang H, Qi Q, Tao F (2022) A multi-scale modeling method for digital twin shop-floor. J
Manuf Syst 62:417–428. https://doi.org/10.1016/j.jmsy.2021.12.011
31. Jones D, Snider C, Nassehi A, Yon J, Hicks B (2020) Characterising the Digital Twin: a
systematic literature review. CIRP J Manuf Sci Technol 29:36–52. https://doi.org/10.1016/j.
cirpj.2020.02.002
32. Morgan CB (1984) Discrete-event system simulation
33. Beck A (2008) Simulation: the practice of model development and use. J Simul 2:67. https://
doi.org/10.1057/palgrave.jos.4250031
34. Yin C, McKay A (2018) Introduction to modeling and simulation techniques
35. Banks J (1998) Principles of simulation. Handbook of simulation: principles, methodology,
advances, applications, and practice, pp 3–30
36. Law AM, Kelton WD, Kelton WD (2007) Simulation modeling and analysis, vol 3. Mcgraw-
Hill New York
37. Schluse M, Rossmann J (2016) From simulation to experimentable digital twins: simulation-
based development and operation of complex technical systems. In: ISSE 2016—2016 inter-
national symposium on systems engineering—Proceedings Papers, Institute of Electrical and
Electronics Engineers Inc. https://doi.org/10.1109/SysEng.2016.7753162
38. McHaney R (1991) Computer simulation: a practical perspective. Academic Press
39. MOSIMTEC (2022) 4 Types of simulation models to leverage in your business 2022. https://
mosimtec.com/types-of-simulation-models/. Accessed 17 Dec 2022
40. Harrison RL (2010) Introduction to monte carlo simulation. In: AIP conference proceedings,
vol 1204. American Institute of Physics, pp 17–21
41. Davidsson P, Holmgren J, Kyhlbäck H, Mengistu D, Persson M (2006) Applications of
agent based simulation. In: International workshop on multi-agent systems and agent-based
simulation. Springer, Heidelberg, pp 15–27
42. Moreira ACM, Silva DF (2021) Applying discrete-event simulation and value stream mapping
to reduce waste in an automotive engine manufacturing plant. Winter Simul Conf (WSC)
2021:1–12. https://doi.org/10.1109/WSC52266.2021.9715477
43. Banks J (1999) Introduction to simulation. In: Proceedings of the 31st conference on Winter
simulation: simulation—a bridge to the future, vol 1, pp 7–13
Contrasting Capabilities: A Comparative Analysis of Simulation … 49

44. White KP, Ingalls RG (2015) Introduction to simulation. Winter Simul Conf (WSC) 2015:1741–
1755. https://doi.org/10.1109/WSC.2015.7408292
45. Ingalls RG (2011) Introduction to simulation. In: Proceedings of the 2011 Winter Simulation
Conference (WSC), pp 1374–88. https://doi.org/10.1109/WSC.2011.6147858
46. Bennett BS (1995) Simulation fundamentals. Prentice Hall International (UK) Ltd
47. Wooley A (2023) Simulation and Digital Twin: contrasting the capabilities and bridging the
gap 2023
48. Chung CA (2003) Simulation modeling handbook: a practical approach. CRC Press
49. Sokolowski JA, Banks CM (2011) Principles of modeling and simulation: a multidisciplinary
approach. John Wiley & Sons
50. Bossel H (2013) Modeling and simulation. Springer, Heidelberg
51. Bandyopadhyay S, Bhattacharya R (2014) Discrete and continuous simulation: theory and
practice. CRC Press
52. Maria A (1997) Introduction to modeling and simulation. In: Proceedings of the 29th conference
on Winter simulation, pp 7–13
53. Shannon RE (1998) Introduction to the art and science of simulation. In: 1998 winter simulation
conference. Proceedings (cat. no. 98ch36274), vol 1, IEEE, pp 7–14
54. Carson JS (2005) Introduction to modeling and simulation. In: Proceedings of the Winter
simulation conference, IEEE, p 8
55. Goldsman D (2007) Introduction to simulation. In: 2007 Winter simulation conference, IEEE,
pp 26–37
56. Sanchez PJ (2006) As simple as possible, but no simpler: a gentle introduction to simulation
modeling. In: Proceedings of the 2006 Winter simulation conference, IEEE, pp 2–10
57. White KP, Ingalls RG (2018) The basics of simulation. In: 2018 Winter Simulation Conference
(WSC), IEEE, pp 147–61
58. Robinson S, Arbez G, Birta LG, Tolk A, Wagner G (2015) Conceptual modeling: definition,
purpose and benefits. Winter Simul Conf (WSC) 2015:2812–2826. https://doi.org/10.1109/
WSC.2015.7408386
59. Smith JS, Sturrock DT (2021) Simio and simulation—modeling, analysis, applications, 6th
edn
60. Davenport TH, Harris JG, Morison R (2010) Analytics at work: smarter decisions, better results.
Harvard Business Press
61. Wright L, Davidson S (2020) How to tell the difference between a model and a digital twin.
Adv Model Simul Eng Sci 7. https://doi.org/10.1186/s40323-020-00147-4
62. Ante L (2021) Digital twin technology for smart manufacturing and industry 4.0: a bibliometric
analysis of the intellectual structure of the research discourse. Manuf Lett 27:96–102. https://
doi.org/10.1016/j.mfglet.2021.01.003
63. Wooley A, Silva DF, Bitencourt J (2023) When is a simulation a digital twin? A systematic
literature review. Manuf Lett 35:940–951
Reinforcement Learning for Digital
Twins

Deena Francis, Jonas Friederich, Adelinde Uhrmacher,

and Sanja Lazarova-Molnar

Abstract Digital Twins (DTs) aim to ongoingly replicate complex systems through
data acquisition, simulation, and analysis to monitor, optimize, and/or experiment to
achieve systems’ goals. Typically, systems adapt and evolve during their lifetimes,
which requires updating simulation models and analysis as new data arrives or con-
ditions change. The dynamics of environments under which DTs commonly operate
necessitate using a paradigm that can deal with the uncertainties and unprecedented
scenarios that may arise throughout its operation. Reinforcement Learning (RL) is
a learning paradigm that provides tools to do precisely this. It is concerned with
sequential decision-making in dynamic, uncertain environments. In this work, we
discuss the current and potential role of RL in the context of DTs, motivate its usage
through a concrete case study, and finally discuss the opportunities and challenges.

1 Introduction

A Digital Twin (DT) is a high-fidelity representation of an intended or current physical

counterpart in terms of a computational model, often in the form of machine learn-
ing models or simulation models [1]. The digital and physical twins are typically

D. Francis (B)
Department of Engineering Technology, Technical University of Denmark,
Ballerup, Denmark
e-mail: [email protected]
J. Friederich
University of Southern Denmark, Odense, Denmark
e-mail: [email protected]
A. Uhrmacher
University of Rostock, Rostock, Germany
e-mail: [email protected]
S. Lazarova-Molnar
Karlsruhe Institute of Technology, Germany and University of Southern Denmark,
Odense, Denmark
e-mail: [email protected]
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 51
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_3
52 D. Francis et al.

connected in near-real-time to exchange information and data, allowing a frequent

“twinning” between the virtual and the physical. DTs are inherently data-driven;
therefore, supervised and unsupervised machine learning plays an important role in
their realization [2]. In addition, due to the twinning, the digital and physical twins
are often updated frequently, bringing us to another class of algorithms that learns
from interacting with their environment, i.e., reinforcement learning. Reinforcement
learning is concerned with sequential decision-making in dynamic, uncertain envi-
ronments.
In the following, we explore how the development and deployment of DTs in a
dynamic environment can be supported by reinforcement learning. Our focus is on
DTs based on discrete event simulation models. In particular, in our case study, we
discuss a DT in which not only parameters of a simulation model are adapted, but
the simulation model itself is learned from data [3]. Discrete decisions are plentiful
within the DT life cycle (as shown in Fig. 1). They may refer to what data to collect,
what methods to use for validating the data, what methods to extract knowledge from
the data, how to develop the model (e.g., which model components to compose), what
methods to use for calibrating or validating the model, and what actions to send back
to the environment. We present several applications of reinforcement learning in the
context of DTs to gain insights about which questions in developing and deploying
DTs can be effectively supported by reinforcement learning.

2 Reinforcement Learning

Reinforcement learning is about sequential decision-making; the agent learns how

to map situations to actions to maximize a cumulative reward from the environment.
Reinforcement learning assumes a closed loop in which the learning system’s actions
will influence its later inputs. The agent discovers which actions yield the most reward
by trial and error. Markov Decision Processes (MDPs) are mathematical models
used to formalize and study sequential decision-making in reinforcement learning
[4]. Table 1 gives a short introduction to reinforcement learning concepts [4]. The
notation used in this chapter is mostly from the third chapter of [4].
Reinforcement learning appears intuitive and, as such, attractive to solve prob-
lems, but its application is not without challenges. One concern is the complexity of
the environment; real-world environments can be highly complex, making it diffi-
cult to model accurately. Generally, RL algorithms struggle when faced with high-
dimensional state and action spaces or when dealing with continuous or stochastic
environments, so it is crucial to generate the state space for RL carefully. The under-
lying hypothesis of applying RL is that the goal can be described as maximizing
the expected cumulative rewards, which might be difficult for some applications.
RL algorithms typically require a large number of interactions with the environment
to learn effective policies. This can be prohibitively expensive, time-consuming,
or impractical in some applications. Exploration and exploitation need to be care-
Reinforcement Learning for Digital Twins 53

Table 1 Central concepts of reinforcement learning

Concept Description
Agents Reinforcement learning concerns at least two active entities, i.e., the
environment and one or more agent(s). An agent receives observation and
reward from the environment and sends actions to the environment; the
agent employs reinforcement learning to maximize its rewards
Environment The environment is also considered an active entity with dynamics of its
own; it receives the actions . A1 , ..., Ai , ... and sends observations
. O1 , ..., Oi , ... and rewards . R1 , ..., Ri , ... to the agent

States An agent uses its state to select its next action, forming the reinforcement
algorithm’s basis. Agent’s state . St is typically a function of the history of
encounters between the agent and its environment, i.e., . St = f (Ht ) with
. Ht = O1 , R1 , A1 , . . . , Ot−1 , Rt−1 , and is assumed to be Markov, i.e., that
the future is independent of the past, given the present. Often, the
environmental state is not fully observable by the agent, which makes it
formally a Partially Observable Markov Decision Process (POMDP)
Actions An agent selects actions to maximize its cumulative reward, .A is the set of
actions an agent selects from
Rewards A reward is a scalar feedback signal that indicates how well the agent is
doing at step .t,
Discount factor The discount factor .γ ∈ [0, 1] influences how much an agent values future
rewards
Value function The action-value function .qπ (s, a) is the expected return starting from
state .s, taking action .a, and following policy .π. The action-value function
can be decomposed into immediate reward and the discounted return from
the successor state,
.qπ (s, a) = Eπ [Rt+1 + γ qπ (St+1 , At+1 )|St = s, At = a]

Model Reinforcement learning approaches are distinguished based on whether

they construct an explicit model of the environment’s dynamics, typically
a = P[S '
in the form of a transition function .Pss ' t+1 = s |St = s, At = a],
or whether they are model-free
Policy A policy .π maps the states to actions and is defined as a distribution over
actions given the states, .π(a|s) = P[At = a|St = s] The goal of RL is to
learn the optimal/near-optimal policy to maximize the action-value
function for all policies .π , .q∗ (s, a) = maxπ qπ (s, a)

fully balanced. Both exploration (trying out new actions to discover their effects)
and exploitation (leveraging known information to maximize rewards) are crucial in
RL. Finding the right balance is challenging, especially in environments with sparse
rewards or deceptive dynamics [4].

3 Application of Reinforcement Learning in Digital Twins

In [5], data collection, data validation, knowledge extraction, model development,

and model validation were identified as constituent parts of DTs. Exemplarily, the
54 D. Francis et al.

Fig. 1 The framework for a DT for a factory, adapted from [5]

Table 2 An example of the use of RL in the data collection of a data-driven DT

Data collection
Main task(s) Recommend strategies to the PT for data collection
Simulation environment The physical twin for a telecommunication network with multiple
physical counterparts like mobile devices and network nodes, which
are the agents. There are . K agents in this environment
Type of algorithm Model-based Bayesian multi-agent reinforcement learning (MARL)
States Network-related artifacts like traffic load conditions, queue length of
packets at the devices denoted as .stk , where .k = 1, 2, ..., K
Actions Decisions related to the network such as send a packet, do not send
the packet denoted as .at
Policy Data collection policy for the PT, .π(atk |h kt ), which assigns a
probability to each action .atk , given the action-observation history
.h t = (o1 , a1 , ..., ot , at , ...). Here .ot is an observable function of the
k k k k k k

state .st such as the traffic load of a base station

Reward Low model uncertainty, which happens when there are only a few
disagreements between the various model predictions
Reference [6]

DT framework in the context of a smart factory is shown in Fig. 1. In this section,

we explore each part of the framework in detail, identifying prospects for how RL
can enhance the related processes. We, furthermore, identify promising examples to
illustrate.

3.1 Data Collection

Data collection is the DT element concerned with real-time data gathering and storage
from the physical counterpart of the DT. This data is used to generate and update
Reinforcement Learning for Digital Twins 55

Table 3 An example of the use of RL in the data validation

Data validation
Main task(s) Determine value of data
Environment Predictor model training and evaluation
Type of algorithm REINFORCE [14] (a neural network-based RL) algorithm adapted for
data value determining problem
States Samples of data
Actions Decision to select data instance (yes or no)
Policy Data value determining policy .π(D, s), where .D is the training dataset
of . N samples, .s ∈ {0, 1} N , a data selection vector
Reward Performance of the predictor on a validation set
Reference [15]

DTs. In the design of DTs, the interest has been centered on the technologies for data
collection, such as Internet of Things (IoT), sensors, etc. [7], but in a DT, there are
other aspects of data collection such as how much and when to collect (frequency)
that are also important.
In order to answer the question of how much data to collect, there are some pre-
vious works that have used pre-training in RL to improve the performance in down-
stream learning tasks while learning the most useful features [8]. It has been found
that when the data is collected passively in finite state episodic Markov Decision Pro-
cesses (MDP), to obtain an .∈-optimal policy, one would need an exponential number
of samples. More formally, to obtain an.∈-optimal policy.Ω(Amin(S−1,H ) ∈ −2 ) episodes
would be needed [9]. The question of when to collect the data can be explored by
looking at an example of RL being employed in the data collection is described in
Table 2. RL has been used in developing optimal data collection strategies in DTs for
wireless communication systems with notable benefits [6]. Such strategies inform
when and what data to collect to refine the RL models.
Their RL-based data collection method was evaluated against other data collection
schemes in an experimental study, and it was found that their method performed
as well as an oracle-aided baseline method as the number of rounds (more data
samples) increased. Gleeson et al. [10] explored the possibility of reducing the data
collection time as well as computation time by utilizing the parallelization offered
by GPUs. They demonstrate a 13.4x faster implementation of a deep RL framework.
Reducing the state space has been done by including sensor selection as part of the
learning task [11]. Long periods of training and/or data collection are often needed to
obtain success in real-world learning problems [12]. Inspired by the success of data
augmentation in supervised learning, [13] proposed using data augmentation for RL
for pixel-based (image) and state-based data. Their results demonstrated that simple
RL methods can outperform state-of-the-art on several tasks. This approach could
possibly benefit RL for DTs in that the training/data collection time can be reduced
while providing performance boosts.
56 D. Francis et al.

Table 4 An example of the use of RL in the knowledge extraction phase of a data-driven DT

Knowledge extraction
Main task(s) Event detection
Environment Battery-less energy harvesting sensors
Type of algorithm Deep RL, Proximal Policy Optimization (PPO)
States Energy storage, time (day, hour, minutes), light intensity
Actions Turn on or off the Passive Infrared (PIR), humidity, temperature, light,
and pressure sensors
Policy Event detection policy .π(a|s)
Reward For every event that it catches, +1 is assigned; for every node that
depletes its energy, .−1 is assigned
Reference [19]

3.2 Data Validation

The data validation of a digital twin consists of tasks, such as imputation, anomaly
detection, etc., which can be solved by supervised and unsupervised learning, respec-
tively [16]. RL has also been used for anomaly detection in sequential and time series
data [17, 18]. In addition, RL can also be used to determine the value of data [15].
This method determines whether the data in hand is useful from a learning/modeling
perspective. Their results indicate that an RL-based approach performed better than
other approaches. Determining the value of data has also been used effectively in
selecting training samples that are similar to the data in a given validation set, which
is quite beneficial in the realm of domain adaptation [15]. The details are shown in
Table 3.
Often, the collected data contain missing values. A strategy to combat this issue is
to perform data imputation. Data imputation has been approached from unsupervised
and supervised perspectives [20]. RL-based approaches for imputing data have also
been studied with some success [21, 22]. Data measured through sensors have mea-
surement noise. Handling noisy data appropriately is a crucial task. [23] proposed
a noise-robust RL method that demonstrated reliable control in a network control
system. Their method consists of an extended Kalman filter with RL.

3.3 Knowledge Extraction

Knowledge extraction is another essential part of the DT framework that involves

event detection and process mining, as well as mining other application- and goal-
relevant elements of the model. The dynamic nature of events in the physical coun-
terpart motivates the use of a learning paradigm such as RL. Event detection is an
important element of DTs, and RL can be effectively used for event detection when
Reinforcement Learning for Digital Twins 57

no historical data is available, as demonstrated by Fraternali et al. [19]. They used

a self-supervised strategy to collect baseline data. An example of RL applied to
knowledge extraction is described in Table 4.

3.4 Model Development and Deployment

In the DT framework outlined in Fig. 1, the model development part is where the
simulation model is generated from the data and extracted knowledge. The DT is
based on this underlying simulation model of the physical counterpart, reflecting its
dynamics and functionalities. The choice of the modeling formalism depends on the
application domain and the goal(s) of the DT. The model’s parameters are updated
at pre-defined intervals to ensure high fidelity. The dynamic nature of the DT makes
adopting an RL model beneficial in this phase. Deployment of the model involves
the model providing insights for decision-making from the PT’s perspective.

3.5 In-situ Learning of Methods to Apply for Digital Twins

Various simulation methods are used in developing, analyzing, and deploying the
digital twin. To make the possible role of RL for selecting and configuring suitable
methods for a DT more concrete, let us look at the selection and configuration of
simulation algorithms. The efficiency of model execution determines how thoroughly
a model can be analyzed and validated. In addition, as the DT is used to inform its
physical twin, the execution efficiency is also decisive for this information to reach
the physical twin in time. Also, the DT needs to reflect changes within its physical
counterpart and shall be functional over a longer period of time; therefore, adaptations
appear as an intrinsic ingredient of the DT. Depending on the simulation model and
the available hardware structure, the efficiency of simulation algorithms varies [24],
implying that for the DT to be effective, simulation algorithms must be selected
and configured automatically. Various approaches have been developed to select and
configure simulation algorithms automatically [25, 26].
The problem of selecting among the different simulators and their configurations,
given the model and the available hardware, can be posed as a learning problem, either
as a supervised learning problem [27] or as a reinforcement learning problem. To
interpret it as a reinforcement learning problem, i.e., choosing iteratively between
different options with previously unknown rewards, we assume that we can use
part of the simulation, e.g., either some part of a long simulation run or a few of
the replications, to identify a setup that performs best with respect to a user-defined
criterion, e.g., simulation speed. The approach to tackle algorithm selection problems
[28] with reinforcement learning is not unusual [29], even to apply it to reinforcement
learning as a meta learner [30]. To illustrate the challenges, we will shortly sketch
the approach developed in [31], which aims at and analyses automatic, frequent
58 D. Francis et al.

Table 5 Reinforcement learning to select the most efficient simulator configurations online [31]
Concept Description
Agents Adaptive simulator
Environment Execution environment, simulation model
States Model information, e.g., variable values or coupling scheme of
components, information about the simulator’s state, e.g., event queue
length, and information about the hardware, e.g., available cores
Actions Set of simulation algorithms (and components that might be configured)
Rewards Rate of event computations per second
Discount factor 0
Value function Q-learning, . Q[s, a] :=
' '
. Q[s, a] + α(N [s, a]) · (r + γ · max a ' Q[s , a ] − Q[s, a])

Model model-free
Policy f(s,Q,N)

adaptations of simulators during simulation runs via reinforcement learning. The

approach is based on Q-learning. Q-values record the value of taking an action
.a ∈ A in a state .s ∈ S. These values are unknown at first but are updated step by
step by applying the following function: . Q[s, a] := . Q[s, a] + α(N [s, a]) · (r + γ ·
maxa ' Q[s ' , a ' ] − Q[s, a]) (Table 5)
Q-learning takes the received reward .r (of applying .a in state .s) and the observed
state .s ' (as a result of taking action .a in state .s) into account to update the q-value.
.γ is the discount factor, which determines the influence of possible future utilities
on the current q-value, and the function .α determines the learning rate, i.e., the
impact an update has on the q-value, which depends on how often the action .a has
already been chosen in a state.s,. N [s, a]. The policy.π : S × R|S|×| A| × N|S|×| A| → A
is responsible for action selection. The policy, selecting the next action, balances
exploitation (based on . Q, choose actions that promise a high utility) and exploration
(based on . N , choose actions that have been rarely chosen in the current state).
In our case, the set of options to select from is defined by the different simulation
algorithms (and components that might be configured) for the models to be executed.
Therefore, a pre-requisite is a clear separation of concern between model and simu-
lator [32] so that simulators can be exchanged and composed even during simulation
runs [33]. A suitable reward (as the goal is to minimize the time needed for a simula-
tion run) can easily be identified and calculated, e.g., the rate of event computations
per second. To define the state space for reinforcement learning, information about
the model state in terms of variable values or coupling scheme of components, infor-
mation about the state of the simulator in terms of event numbers per minute or event
queue length, and information about the environment in terms of available cores
might be valuable. The discount factor .γ was set to zero, as experiments had shown
that, in the benchmark cases used, a higher value of.γ had no significant impact on the
results [31]. This may be different when using stochastic simulation algorithms and
Reinforcement Learning for Digital Twins 59

data structures with longer warm-up phases. With .γ = 0, the reinforcement learning
problem reduces to searching (i.e., it is assumed that actions do not affect future
situations and rewards). To serve as a basis of reinforcement learning, this complex,
highly-dimensional information needs to be suitably aggregated into more simple
states. Therefore, dynamic, flexible approaches are required [34], which might ben-
efit from applying learning methods [35]. In any case, a suitable partitioning and
aggregation of the state space is essential for an effective reinforcement learning of
simulator selection and configuration [31].
Also, the question of when to trigger a reinforcement learning step needs to be
answered: it could depend simply on the time passed since the last learning step or
be triggered by drastic changes in the dynamics. To identify such changes, informa-
tion about the current and previous states of the model, simulator, or hardware and
the current performance must be interpreted to identify suitable points for adapt-
ing (respectively when invoking the reinforcement step) [36]. Generally, change
point detection [37] is an important ingredient to support decisions when dealing
with dynamic systems. Due to the frequent feedback (performance data) between
the agent (adaptive simulator) and environment (execution environment including
model, chosen simulator configuration, hardware), suitable simulator configurations
can be learned effectively [31] and facilitate leveraging algorithms and data structures
tailored to specific requirements and application contexts.
One could apply reinforcement learning to support other tasks in developing the
DT by method selection and configuration. For example, reinforcement learning
has also been proposed to learn optimization algorithms [38, 39]. The approach
uses guided policy search at the meta-level to train the optimizers and showed good
performance in optimizing neural networks.
The agent and environment must frequently interact for reinforcement learning
to be effective. Therefore, sample efficiency is critical for the in-situ learning of
methods in the context of DT. Other conditions are that an immediate reward can be
easily calculated and is due to the action (credit assignment problem). Also, it should
be noted that reinforcement learning is only one possibility to solve the algorithm
selection problem [40], even in the context of machine learning approaches, and thus
to support developing or deploying the DT effectively.

4 Case Study in Reliability-Focused Digital Twins

of Production Systems

In this section, we present an illustrative case study of a reliability-focused DT to

demonstrate the application potential of RL in DTs. A reliability-focused DT aims to
support decision-making processes related to equipment purchases (i.e., whether to
invest in a more reliable piece of equipment), maintenance scheduling and staffing,
and other system configuration decisions that impact the systems’ reliability. When
considering the life cycle of a DT in the context of a smart factory in Fig. 1, we focus
60 D. Francis et al.

AGV 1 Cell 1

Transport to assembly cell 1 Assembly operation

MES MES
OR

Initiate new production order Complete production order

AGV 2 Cell 2

Transport to assembly cell 2 Assembly operation

Fig. 2 Overview of the case study system

solely on the knowledge extraction, model development, and model validation phase
in this case study.
The following subsections describe the case study system, followed by the DT
extraction procedure. We then examine how an extracted DT model is validated.
Finally, we reflect and discuss the different application areas of RL for this case
study.

4.1 Case Study System: A Flow Production Line

with Concurrency

Figure 2 provides an overview of the case study system, which is a flow produc-
tion line commonly found in manufacturing systems. The production line is fully
automated and consists of five resource components: a manufacturing execution sys-
tem (MES), two automated guided vehicles (AGVs), and two assembly cells. Both
assembly cells work concurrently, performing the same assembly operation. The
MES controls the production process by initiating new production orders, routing
them to either assembly cell 1 or 2, and marking orders as completed. When a new
production order is initiated and assigned to one of the assembly cells, the AGVs
transport the raw material to the designated cell.
The AGVs and the assembly cells are susceptible to failures, while the MES is
always fully operational. In the event of a production resource failure, the resource
stops operating, and a repair crew is dispatched to repair the malfunctioning resource.
The maintenance policy in place is purely reactive. The AGV has an unlimited buffer
and a capacity of one, while both assembly cells 1 and 2 have finite buffers and
capacities of one.
We assume that several types of data can be collected from the case study system,
including event logs, which capture production processes, and state logs, which
record operational state changes of production resources. These two log types are
necessary for the DT extraction process, described in the next section.
Reinforcement Learning for Digital Twins 61

4.2 Digital Twin Model Extraction for the Case Study System

To extract a reliability-focused DT of the case study system, we employ our method-

ology described in [3]. Our proposed methodology uses advanced process mining
(PM) and statistical techniques to extract and parameterize a stochastic Petri net
(SPN), which serves as a conceptual model for the DT of the physical system. In the
following, we describe the extraction and parameterization process in further detail
and present the resulting DT for the case study system.
The class of SPNs that we use to describe the conceptual model for a DT of a
physical system is defined as follows:

. S P N = (P, T, A, m 0 )

where:
• . P = {P1 , P2 , .., Pp } is the set of places, drawn as circles,
• .T = {T1 , T2 , .., Tq } is the set of transitions along with their distribution functions
or weights, drawn as bars,
• . A = A I ∪ A O ∪ A H is the set of arcs, where . A O is the set of output arcs, . A I is
the set of input arcs, and . A H is the set of inhibitor arcs, and each of the arcs has a
multiplicity assigned to it,
• .m 0 is the initial marking of the Petri net.
Each transition .Ti can be either timed or immediate. A timed transition is drawn
as a hollow bar and is assigned a probability distribution function that describes
the firing time of the corresponding activity. An immediate transition is drawn as
a filled-out bar and is assigned a weight that describes the firing probability of the
associated event.
To extract the reliability-focused DT, we first extract a manufacturing process
model and then integrate it with fault models for production resources.
The extraction of the manufacturing process model utilizes the event log and
involves the following four steps:

1. Identification of a Petri net of the material flow within a production line using
process discovery algorithms. The material flow is the path that production orders
follow through the system [41].
2. Determination of transition types (i.e., timed or immediate). Timed transitions
correspond to the arrival of new production orders or to resource activities.
3. Estimation of probability distributions for timed transitions and extraction of
weights for immediate transitions.
4. Extraction of resource capacities and buffers.
62 D. Francis et al.

direct_to_line_1 transport_to_cell_1_buffer enter_cell_1 cell1_operation

mes agv1 mes cell1
0.6 normal(4,1) 1 normal(7,1)
5 repair
lognormal(4,0.2)

cell1 cell1 OK
failed
fail
agv2 failed weibull(20,90)
fail repair order completed
weibull(20,100) lognormal(5,0.2) mes
1
new_order agv1 OK agv1 failed agv2 OK
mes
exponential(0.3) repair fail
lognormal(5,0.2) weibull(20,100)
5 repair
lognormal(3,0.2)

direct_to_line_2 transport_to_cell_2_buffer enter_cell_2 cell2_operation cell2 cell2 OK

mes agv2 mes cell2 failed
0.4 normal(6,1) 1 normal(8,1)
fail
weibull(20,80)

Fig. 3 Extracted DT of the case study system

The extraction of the fault models for production resources utilizes the information
captured by the state log and involves the following two steps:

1. Creation of necessary places and transitions from a fault model template.

2. Estimation of resource failure and repair distributions.

Finally, the fault models are integrated into the manufacturing process model
using inhibitor arcs, preventing the corresponding resource activity timed transition
from firing.
Figure 3 depicts the extracted DT obtained through the described model extraction
process. The figure shows the extracted and parameterized manufacturing process
model as well as the resource fault models. For each timed transition, the corre-
sponding distribution function, including parameters, is shown. For each immediate
transition, the corresponding weight is displayed. Furthermore, the capacity of one
for both assembly cells, as well as their finite buffer sizes, have also been extracted.

4.3 Validation of Extracted Digital Twins

Validation is an essential step to ensure that the extracted DT accurately mimics the
real-world system. In [42], we propose to validate extracted models in two phases:
validation of initial model and validation of model at run-time. In the first phase, the
validity of a newly extracted DT model is evaluated to ensure that the model is safe
to deploy in a production environment to support decisions. In the second phase, the
deployed DT model is validated to ensure continuous validity.
Reinforcement Learning for Digital Twins 63

Table 6 Applying RL for data collection

Concept Description
Goal Determine data streams and amounts of data needed for model extraction
Environment Cyber-physical system with numerous IoT devices
States Set of data streams and sizes of data considered for extraction, structure of
the PN, including model validation results
Actions Adding or removing data streams, Extending and reducing size of data,
Decreasing and increasing frequency of data
Rewards Model validation results for the extracted PN

4.4 Application of Reinforcement Learning to the Case Study

In the following, we investigate and illustrate the application of RL to the case study,
considering all phases of the DT life cycle. Furthermore, we highlight how the case
study system can be optimized using RL and the extracted reliability-focused DT.

Data Collection

The data collection process for reliability-focused DTs can be supported by using
RL to determine the right data streams and the amounts of data needed for model
extraction (Table 6). By adjusting data streams, their size, and frequency of collec-
tion, the impact on the model validation results of the extracted PN model can be
observed and fed back to the agent as a reward. The main challenge in this phase is
the possible large size of the state space, which can lead to high time complexity [43].

Data Validation

RL can be used to to learn and identify patterns, anomalies, or errors within datasets
such as event logs and state logs used for model extraction (Table 7). To do so, spe-
cific metrics or criteria that define the quality and accuracy of the dataset need to be
established. These could include measures such as data consistency, completeness,
accuracy, and adherence to defined standards. The data value determination method
of [15] could serve as a possible solution to selecting useful or error-free data.

Model Extraction

One possible application of RL in the DT model extraction phase could be for material
flow identification (Table 8). RL can be used to interactively extract the material
flow model using process discovery algorithms where the system learns from user
feedback. For instance, when a user corrects or validates certain parts of the material
flow model, the RL system can learn which types of structures are preferred or which
common mistakes to avoid in the following process discovery iterations.
64 D. Francis et al.

Table 7 Applying RL for data validation

Concept Description
Goal Identify and address anomalies, errors, and inconsistencies within a dataset
Environment The dataset
States Subsets of data, specific data points, or statistical summaries of the dataset
Actions Flagging data points as potential anomalies, suggesting corrections or
imputations for missing values, or triggering automated data cleaning
processes
Rewards Positive rewards for improvements in data quality and negative rewards
for data quality degradation

Table 8 Applying RL for the material flow identification

Concept Description
Goal Extract an accurate material flow model
Environment Material flow Petri net (PN) model
States Structure of the PN
Actions Adding or removing places, transitions, and arcs
Rewards User validation of correct and incorrect PN modeling elements

Similarly, RL can be applied to improve the final extracted DT model interactively.

Here, the actions would not just include adding or removing places, transitions, and
arcs but also adjusting the token distribution, transition types, arc multiplicity, distri-
bution functions, and weights. A possible issue is the requirement for a large number
of steps in order to find the optimal policy. In one study, several deep RL methods
were employed for the task of optimization in a manufacturing setup with some suc-
cess. Their problem had a state vector with 12 variables and three possible actions per
state. In this setting, the Proximal Policy Optimization (PPO) algorithm was found
to converge toward the optimal policy in around 160k steps.

Model Validation

RL can support the model validation process of an extracted model by determining

the right validation strategy as well as the time intervals at which validation should
take place (Table 9). For the latter, an RL agent could adjust the validation intervals
and receive a reward based on the validation results. The work of [44] is an example
of RL being used to successfully validate a model in three varied engineering systems.

System optimization

RL can be used in several ways to optimize the real system by making adjustment
to the DT and feeding the gained knowledge back to the real system (Table 10). This
Reinforcement Learning for Digital Twins 65

Table 9 Applying RL for model validation

Concept Description
Goal Determine the right time intervals for model validation
Environment The extracted DT and the physical system it mimics
States Validation intervals
Actions Adjusting validation intervals
Rewards Validation results

Table 10 Applying RL for joint model and system optimization

Concept Description
Goal Enable automated decision support for the real system
Environment The extracted DT and the real system
States Production process states
Actions Adjustments to Buffer sizes, logistic routes, schedules, etc.
Rewards KPI changes

could eventually lead to automated decision support in the real system, unleashing
the full potential of RL. Some of the activities RL could help optimize the real system
are:
• Find optimal parameters to increase/improve a KPI (e.g., production volume,
resource downtime).
• Use an RL agent to determine routing decisions instead of hard-coded routing
probabilities (i.e., direct_to_line_1 and direct_to_line_2 transitions).
• Order scheduling: Use an RL agent to determine production order schedule, instead
of using manually defined order schedules (i.e., new_order transition).
• Buffer size allocation: Use an RL agent to dynamically adjust buffer sizes based
on system state (i.e., buffer for cell 1 and cell 2).
• Optimal repair scheduling: Use RL agent to schedule and dispatch repair crews.

A deep Q-network model was used in the regulation of a subsystem of the Fermilab
booster accelerator complex with notable success in a test scenario [45][46]. RL has
also been demonstrated to be effective model optimization in a sheet metal assembly
setup [47].

5 Discussion and Conclusion

In our study, we investigated how Digital Twins can be enhanced using Reinforcement
Learning at the different steps of Digital Twins’ lifecycle. Reinforcement Learning
is suited to the scenarios found in Digital Twins where sequential decision-making
66 D. Francis et al.

needs to be done based on interactions with an environment that is dynamic. We,

furthermore, provided a case study to illustrate the different opportunities. There
are several aspects of utilizing Reinforcement Learning in Digital Twins that lead to
opportunities and challenges, which we briefly describe in the following.
• Quality of data: Large quantities of data are generated from the physical counter-
part of the Digital Twin setup, which aids the Reinforcement Learning model to
learn optimal policies. The downside is, however, that not all data that has been
collected is relevant to building a reliable model. Determining the quality or value
of data has been studied by [15] by formulating this problem as an RL problem.
• Curse of dimensionality: When the state-action space is large, reinforcement learn-
ing methods will suffer from high time complexity. The Digital Twin scenario will
typically have a large state-action space as discussed in the case study (Sect. 4).
Several works have attempted to tackle this issue from various perspectives [48,
49].
• Pre-training: In systems with sensors that collect videos or audio signals for long
periods of time, large amounts of data are available. It has been found that using
this data to learn a good representation of the state improves the performance of
the downstream Reinforcement Learning task [8]. It would be interesting to see if
such a strategy would also benefit a Digital Twin scenario.
• Uncertainty quantification: The Digital Twin model needs to be calibrated to
account for the uncertainties with respect to what it knows about its physical
counterpart. Failure to do so would result in the physical counterpart receiving
poor decisions, which can lead to failures. Quantifying of the uncertainty can be
done using a Bayesian perspective [6].
To summarize, we can state that there remains much to be done and investigated in
the application of Reinforcement Learning in Digital Twins. Our initial investigation
provides initial directions and opportunities that are worth further investigation.

References

1. Grieves M, Vickers J (2017) Digital twin: Mitigating unpredictable, undesirable emergent

behavior in complex systems. Transdisciplinary Perspect Complex Syst New Find Approaches
85–113
2. Huang Z, Shen Y, Li J, Fey M, Brecher C (2021) A survey on ai-driven digital twins in industry
4.0: smart manufacturing and advanced robotics. Sensors 21(19):6340
3. Friederich J, Lazarova-Molnar S (2022) Data-driven reliability modeling of smart manufactur-
ing systems using process mining. In: 2022 winter simulation conference (WSC) pp 2534–2545,
Dec 2022. https://doi.org/10.1109/WSC57314.2022.10015301. ISSN: 1558-4305
4. Sutton RS, Barto AG (2018) Reinforcement learning: an introduction. MIT Press
5. Friederich J, Francis DP, Lazarova-Molnar S, Mohamed N (2022) A framework for data-driven
digital twins of smart manufacturing systems. Comput Ind 136:103586
6. Ruah C, Simeone O, Al-Hashimi B (2023) A bayesian framework for digital twin-based control,
monitoring, and data collection in wireless systems. IEEE J Sel Areas Commun
Reinforcement Learning for Digital Twins 67

7. Liu Mengnan, Fang Shuiliang, Dong Huiyue, Cunzhi Xu (2021) Review of digital twin about
concepts, technologies, and industrial applications. J Manufact Syst 58:346–361
8. Ghosh D, Bhateja CA, Levine S (2023) Reinforcement learning from passive data via latent
intentions. Int Conf Mach Learn 11321–11339. PMLR
9. Xiao C, Lee I, Dai B, Schuurmans D, Szepesvari C (2022) The curse of passive data collec-
tion in batch reinforcement learning. In: International conference on artificial intelligence and
statistics. PMLR, pp 8413–8438
10. Gleeson J, Snider D, Yang Y, Gabel M, de Lara E, Pekhimenko G (2022) Optimizing data
collection in deep reinforcement learning. arXiv preprint arXiv:2207.07736
11. Kishima Yasutaka, Kurashige Kentarou (2013) Reduction of state space in reinforcement learn-
ing by sensor selection. Artif Life Rob 18:7–14
12. Kalashnikov D, Irpan A, Pastor P, Ibarz J, Herzog A, Jang E, Quillen D, Holly E, Kalakrishnan
M, Vanhoucke V et al (2018) Qt-opt: Scalable deep reinforcement learning for vision-based
robotic manipulation. arXiv preprint arXiv:1806.10293
13. Laskin Misha, Lee Kimin, Stooke Adam, Pinto Lerrel, Abbeel Pieter, Srinivas Aravind (2020)
Reinforcement learning with augmented data. Adv Neural Inf Process Syst 33:19884–19895
14. Williams RJ (1992) Simple statistical gradient-following algorithms for connectionist rein-
forcement learning. Mach Learn 8:229–256
15. Yoon J, Arik S, Pfister T (2020) Data valuation using reinforcement learning. Int Conf Mach
Learn 10842–10851. PMLR
16. Chandola Varun, Banerjee Arindam, Kumar Vipin (2009) Anomaly detection: a survey. ACM
Comput Surv (CSUR) 41(3):1–58
17. Oh M-H, Iyengar G (2019) Sequential anomaly detection using inverse reinforcement learning.
In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery
and data mining, pp 1480–1490
18. Mengran Yu, Sun Shiliang (2020) Policy-based reinforcement learning for time series anomaly
detection. Eng Appl Artif Intell 95:103919
19. Fraternali F, Balaji B, Sengupta D, Hong D, Gupta RK (2020) Ember: energy management of
batteryless event detection sensors with deep reinforcement learning. In: Proceedings of the
18th conference on embedded networked sensor systems, pp 503–516
20. Emmanuel Tlamelo, Maupong Thabiso, Mpoeleng Dimane, Semong Thabo, Mphago Banyat-
sang, Tabona Oteng (2021) A survey on missing data in machine learning. J Big Data 8(1):1–37
21. Mei H, Li J, Shi B, Wei H (2023) Reinforcement learning approaches for traffic signal control
under missing data. arXiv preprint arXiv:2304.10722
22. Yamaguchi N, Fukuda O, Okumura H (2020) Model-based reinforcement learning with miss-
ing data. In: 2020 eighth international symposium on computing and networking workshops
(CANDARW). IEEE, pp 168–171
23. Bui V-P, Raj Pandey S, Ana PM, Popovski P (2023) Value-based reinforcement learning for
digital twins in cloud computing. arXiv preprint arXiv:2311.15985
24. Jeschke M, Ewald R (2008) Large-scale design space exploration of ssa. In: Computational
methods in systems biology: 6th international conference CMSB 2008, Rostock, Germany,
October 12-15, 2008. Springer, Berlin, pp 211–230
25. Ewald R (2012) Automatic algorithm selection for complex simulation problems.
Vieweg+Teubner
26. Gratl FA, Seckler S, Tchipev N, Bungartz H-J, Neumann P (2019) Autopas: Auto-tuning for
particle simulations. In 2019 IEEE international parallel and distributed processing symposium
workshops (IPDPSW). IEEE, pp 748–757
27. Ewald R, Himmelspach J, Uhrmacher AM (2008) An algorithm selection approach for simu-
lation systems. In: 2008 22nd workshop on principles of advanced and distributed simulation.
IEEE, pp 91–98
28. Rice JR (1976) The algorithm selection problem. In: Advances in computers, Vol. 15, pp
65–118. Elsevier
29. Lagoudakis MG, Littman ML et al (2000) Algorithm selection using reinforcement learning.
In: ICML, pp 511–518
68 D. Francis et al.

30. Afshar RR, Zhang Y, Vanschoren J, Kaymak U (2022) Automated reinforcement learning: An
overview. arXiv preprint arXiv:2201.05000
31. Helms T, Ewald R, Rybacki S, Uhrmacher AM (2015a) Automatic runtime adaptation for
component-based simulation algorithms. ACM Trans Model Comput Simul (TOMACS)
26(1):1–24
32. Zeigler BP, Praehofer H, Kim TG (2000) Theory of modeling and simulation. Academic Press
33. Himmelspach J, Uhrmacher AM (2007) Plug’n simulate. In: 40th annual simulation symposium
(ANSS’07). IEEE, pp 137–143
34. Helms T, Mentel S, Uhrmacher A (2016) Dynamic state space partitioning for adaptive sim-
ulation algorithms. In: Proceedings of the 9th EAI international conference on performance
evaluation methodologies and tools, pp 149–152
35. Asadi K, Abel D, Littman ML (2020) Learning state abstractions for transfer in continuous
control. arXiv preprint arXiv:2002.05518
36. Helms T, Reinhardt O, Uhrmacher AM (2015b) Bayesian changepoint detection for generic
adaptive simulation algorithms. In: Proceedings of the 48th annual simulation symposium, pp
62–69
37. Aminikhanghahi S, Cook DJ (2017) A survey of methods for time series change point detection.
Knowl Inf Syst 51(2):339–367
38. Li K, Malik J (2016) Learning to optimize. arXiv preprint arXiv:1606.01885
39. Chen T, Chen X, Chen W, Heaton H, Liu J, Wang Z, Yin W (2022) Learning to optimize:
a primer and a benchmark. J Mach Learn Res 23(189):1–59. http://jmlr.org/papers/v23/21-
0308.html
40. Kerschke P, Hoos HH, Neumann F, Trautmann H (2019) Automated algorithm selection: Survey
and perspectives. Evol Comput 27(1):3–45
41. Lugaresi G, Matta A (2021) Automated manufacturing system discovery and digital twin
generation. J Manufac Syst 59:51–66, Apr 2021. ISSN 0278-612 https://doi.org/10.1016/j.
jmsy.2021.01.005. https://www.sciencedirect.com/science/article/pii/S0278612521000054
42. Friederich J, Lazarova-Molnar S (2023) A framework for validating data-driven discrete-event
simulation models of cyber-physical production systems. In: Proceedings of the 2023 winter
simulation conference. San Antonio, USA, IEEE
43. Dietterich TG (2000) An overview of maxq hierarchical reinforcement learning. In: Inter-
national symposium on abstraction, reformulation, and approximation. Springer, Berlin, pp
26–44
44. Schena L, Marques P, Poletti R, Ahizi S, Van den Berghe J, Mendez MA (2023) Reinforce-
ment twinning: from digital twins to model-based reinforcement learning. arXiv preprint
arXiv:2311.03628
45. Kafkes D, Schram M (2021) Developing robust digital twins and reinforcement learning for
accelerator control systems at the fermilab booster. arXiv preprint arXiv:2105.12847
46. John JS, Herwig C, Kafkes D, Mitrevski J, Pellico WA, Perdue GN, Quintero-Parra A, Schup-
bach BA, Seiya K, Tran N et al (2021) Real-time artificial intelligence for accelerator control:
A study at the fermilab booster. Phys Rev Accelerators Beams 24(10):104601
47. Cronrath C, Aderiani AR, Lennartson B (2019) Enhancing digital twins through reinforcement
learning. In: 2019 IEEE 15th international conference on automation science and engineering
(CASE). IEEE, pp 293–298
48. Yang L, Wang M (2020) Reinforcement learning in feature space: matrix bandit, kernels, and
regret bound. In: International conference on machine learning. PMLR, pp 10746–10756
49. Osband I, Van Roy B (2014) Near-optimal reinforcement learning in factored mdps. Adv Neural
Inf Process Syst 27
A Framework for the Credibility
Evaluation of Digital Twins

Han Lu, Lin Zhang, Kunyu Wang, Zejun Huang, Hongbo Cheng,
and Jin Cui

Abstract The widespread applications of Digital Twins are facing a significant chal-
lenge due to the lack of systematic and effective credibility evaluation methods. To
address this issue, this chapter analyzes the dynamic evolution, virtual-real interac-
tivity, and other key features of Digital Twins. Subsequently, a credibility evaluation
framework for Digital Twins is proposed, encompassing the concept of Digital Twins
credibility, a multi-dimensional and multi-level credibility evaluation index system,
and a credibility evaluation methodology. The evaluation process is elucidated using
the robotic arm Digital Twin as an example, thereby providing valuable guidance for
the evaluation and construction of Digital Twins.

Keywords Digital twin · Credibility evaluation · Dynamic evolution · Index

system · Modeling and simulation

1 Introduction

A Digital Twin can be taken as a digital model of a physical object that under-
goes real-time self-evolution through data integration from its physical counterpart,
ensuring consistent synchronization throughout its entire lifecycle [1]. Simulations
based on the model can help to optimize and make decisions on the physical object
[1, 2]. The physical object can be any type of physical system, such as a transporta-
tion system, a factory, a production line, or an equipment. The Digital Twins in
this chapter mainly focus on equipment. The equipment can be the product of the
manufacturing industry or the equipment used in the manufacturing process, e.g.,

H. Lu · L. Zhang (B) · K. Wang · Z. Huang · H. Cheng · J. Cui

Beihang University, Beijing 100191, China
e-mail: [email protected]
Engineering Research Center of Complex Product Advanced Manufacturing System, Ministry of
Education, Beijing 100191, China
State Key Laboratory of Intelligent Manufacturing System Technology, Beijing 100191, China
L. Zhang
Hangzhou International Innovation Institute of Beihang University, Hangzhou 311115, China

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 69

M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_4
70 H. Lu et al.

the aircraft, car, machine tool, or robot. The adoption of Digital Twins in the equip-
ment’s lifecycle offers numerous benefits, including enhanced equipment innovation
and design, reduced development cycles, cost reductions, and improved autonomy in
equipment operation and maintenance [3–8]. During the research and design stage
of equipment development, the utilization of modeling and simulation-based system
engineering techniques enables activities to be conducted within the digital realm [9,
10]. Consequently, design schemes can be thoroughly tested and pre-demonstrated,
leading to significant reductions in physical experiment investments.
Credibility is a metric to measure the “correctness” of a model and the model’s
results for a specific use or purpose [11], which is the degree of user’s confidence in the
correctness of a model [8]. The concept of Verification, Validation, and Accreditation
(VV&A) was proposed by the US Department of Defense (DoD) [12] to guarantee
the credibility of modeling and simulation [13]. How to evaluate the credibility of a
model is a critical task and foundation of VV&A. As a self-evolving and dynamic
model, Digital Twin’s credibility is much more difficult to be evaluated. In practice,
lacking of credibility evaluation for Digital Twins is one of the critical bottlenecks
in the widespread utilization of Digital Twins, as well as Digital Twins. Only a
credible Digital Twin can effectively capture the equipment’s characteristics and
status in a timely manner, thereby assisting decision-makers in making accurate
judgments. Considering that the Digital Twin maintains real-time data interaction
with its physical counterpart, the usage of an untrustworthy Digital Twin may lead
to failed tasks and even catastrophic consequences.
The research on credibility evaluation of Digital Twins is currently limited world-
wide. Consequently, the widespread adoption and extensive application of Digital
Twins face significant barriers. Existing credibility evaluation methods for Digital
Twins primarily focus on visual resemblance and simulation output results, which
are insufficiently comprehensive and scientific for complex Digital Twins.
Firstly, it is essential to recognize that a digital model cannot perfectly repli-
cate a physical object. The excessive pursuit of realism or high accuracy may result
in unnecessary complexity and introduce additional uncertainty, thereby reducing
model credibility and usability. It is crucial to strike a balance between accuracy
and complexity to facilitate efficient model computation and utilization. Secondly,
evaluating credibility solely based on the consistency of output data collected at a
certain time interval between the Digital Twin and its physical counterpart is insuf-
ficient. While external performance similarity is valuable, it does not guarantee that
the internal mechanistic structure of the Digital Twin matches that of the physical
object. Thirdly, relying on data samples from a specific time period fails to represent
all the potential output variations of the model. Therefore, the credibility of Digital
Twins cannot be solely determined by the visual resemblance or simulation output
results.
In summary, the credibility evaluation of Digital Twins demands a comprehensive
and systematic approach. There is an urgent need for the development of a compre-
hensive evaluation theory and methodology. Such methodologies can offer practical
solutions for assessing the credibility of Digital Twins. Moreover, they can provide
guidance on constructing credible Digital Twins, thus ensuring their credibility for
various applications.
A Framework for the Credibility Evaluation of Digital Twins 71

2 Related Work

During the past decade of concept development related to DTs, most of the litera-
ture focused on discussing the concept itself and its underlying connotations [14–
17]. However, there remains a notable dearth of a systematic credibility evaluation
framework for reference within the field. Evaluating a DT entails assessing both the
DT model itself and the real-time data utilized for its construction and evolution.
While theoretically, data changes should be incorporated and reflected in the model,
a separate evaluation of the credibility of the data can help prevent the introduc-
tion of untrustworthy data into the DT, thereby enhancing DT credibility. Despite
the limited resources specifically addressing credibility evaluation for DTs, there is
considerable research available in the field of M&S that may serve as a valuable
resource to develop credibility evaluation methods for DTs.

2.1 Data Credibility Evaluation

The credibility of data used in a DT significantly influences the credibility of the DT.
These data encompass a large number of sensor data and non-sensor data related to
human activities.
In the case of sensor data, the reliability can be compromised if the sensor is
damaged or experiences failure [18]. Additionally, if a sensor fails to function prop-
erly within a sensor network, the data from that sensor node may be considered
unreliable [19]. Even under normal circumstances, sensor data can be influenced
by various factors such as the working environment [20], as well as the installation
distance and angle of the sensors [21]. To evaluate the credibility of sensor data itself,
Liao et al. [22] utilized a cumulative residual Chi-square check to compare current
data with historical data and determine their credibility.
Digital Twin involves human participation during both the construction and oper-
ation decision-making stages. The credibility of the data in Digital Twin systems has
been evaluated from various perspectives. For instance, the influence of ambiguity
and authority on data credibility has been studied through psychological evaluations
[23]. Additionally, the credibility of human-related propositions has been assessed
based on a priori knowledge and a posteriori knowledge [24].

2.2 Model Credibility Evaluation

The existing methods for model credibility evaluation mainly come from the
M&S. The specific methods for assessing model credibility can be categorized
into three groups: qualitative analysis methods, quantitative analysis methods, and
comprehensive analysis methods.
72 H. Lu et al.

As an important means of ensuring credibility, VV&A-related technologies has

been widely studied. Research on VV&A of weapon equipment system simulation
has been carried out by the research team on system simulation at the National
University of Defense Technology [25, 26]. Additionally, the teams at Harbin Institute
of Technology have achieved significant progress in VV&A and credibility evaluation
research on distributed interactive simulation systems [27]. In the field of aerospace
simulation systems, Kim et al. [28] have proposed a trust evaluation framework for
the entire life cycle of modeling and simulation development. The trust evaluation
theory of complex weapon equipment models has been discussed and elucidated in
references by Zhou et al. [29], Zhang and Ye [30], Zhen and Hu [31], and Sim and
Lee [32].
More specifically, Beydoun et al. [33] proposed a set of model evaluation methods
based on expert scoring, simulation requirements, and the simulation environment.
Acar [34] suggested that the predictive ability of meta-modeling can be enhanced by
integrating various types of models using weighted average integration. Li et al. [35]
investigated the credibility evaluation technology of complex simulation models
using the multi-agent interactive network method. Addressing the issue of high
complexity in simulation models, Ferson and Oberkampf [36] developed the u-
pooling region indicator. Li et al. [37] introduced the multi-variate probability integral
transformation (PIT). Dornheim and Brazauska [38] proposed a hybrid linear expec-
tation model for determining the credibility of complex systems in an automated and
efficient manner. Liang et al. [39] put forward a credibility measurement method
based on dynamic Bayesian networks. Hu [40] proposed a dynamic data-driven
simulation method framework that utilizes Monte Carlo simulation to model real-
time wildfire scenarios. Wang et al. [41] devised an incremental external attention
temporal convolution network (IExATCN) model to establish a dynamic evolution
framework for black-box Digital Twins, efficiently balancing computation accuracy
and efficiency. Wang et al. [42] developed a lifelong learning method based on event-
triggered online frozen-Elastic Weight Consolidation (EWC) transformer encoder for
the dynamic evolution of Digital Twins, effectively managing computation costs and
dynamic evolution performance.
However, traditional M&S primarily focuses on offline simulations. DT, on the
other hand, is specifically designed for online simulations, incorporating interactive
and evolutionary characteristics based on traditional offline simulations. Therefore,
existing credibility concepts and evaluation methods from M&S could serve as refer-
ences. However, considering the unique features of Digital Twins, new principles and
methods must be proposed to implement effective credibility evaluation of DTs.

3 Concept of Digital Twin Credibility

The current evaluation of credibility for complex simulation systems primarily

focuses on correctness based on user requirements. However, the analysis for dynam-
ically changing models is often overlooked or not given enough emphasis. As
A Framework for the Credibility Evaluation of Digital Twins 73

a result, existing methods cannot be directly applied to assess the credibility of

Digital Twins. Digital Twins are highly interactive and evolve in real time, making
it crucial to examine whether they can provide correct simulation results in a timely
manner during interaction and evolution. Furthermore, the credibility requirements
for equipment vary greatly throughout its life cycle, leading to different focal points
in credibility evaluation.
The primary purpose of credibility evaluation is to thoroughly analyze the quality
of simulation results from the perspective of the model user’s interests. This allows
users to gain a full understanding of the limitations of the results and their impact
on the decision-making process. As the understanding of credibility and evaluation
methods is still in the exploratory stage, the quantified credibility values may be
somewhat incomplete. Therefore, it is necessary to include a description of this
incompleteness in the credibility analysis to provide decision-makers with a more
comprehensive understanding of the credibility of Digital Twins.
The credibility value serves as a quantitative measure of credibility. In this
chapter, the concept of credibility for Digital Twins is defined as follows: Based
on specific user requirements, and during continuous virtual-real interactions and
dynamic evolutions, the credibility of a Digital Twin refers to the correctness and
timeliness of the twin model, the evolution process, and the simulation results, as
well as the uncertainty analysis in each stage of the entire life cycle.
Figure 1 illustrates the process of assessing the credibility of Digital Twins based
on user requirements. A Digital Twin can be considered credible when the difference
between its behavior and that of the physical object falls within the tolerance of the
user’s requirements.
The fulfillment of user requirements is measured through two main aspects:
correctness and timeliness. Correctness and timeliness are assessed by evaluating
the mapping model between the Digital Twin and physical objects, the evolution
process of the Digital Twin, and the simulation results generated using the Digital
Twin. Timeliness refers to the ability to provide the necessary simulation results
within an acceptable time period upon request. The specific evaluation elements for
correctness and timeliness are defined in the index system presented in Sect. 5.

Fig. 1 Credibility of DT
74 H. Lu et al.

Uncertainty analysis plays a crucial role in credibility evaluation. Uncertainty

exists in various aspects of deriving quantitative values for correctness and timeliness.
The evaluation methodology outlined below provides a methodology for analyzing
different types of known uncertainties.
The parameters, structure, and mechanisms of the twin model undergo changes
and evolution over time. This ongoing process of assessing the credibility of Digital
Twins distinguishes it from the evaluation of traditional basic models, which solely
focus on providing simulations without considering real-time data and evolution
mechanisms.

4 Features of DT Credibility Evaluation

The credibility evaluation of DT is a complex systematic process. The model for

DT is profoundly intricate owing to the multi-disciplinary and multi-level attributes
of the equipment itself. It possesses two fundamental elements of a Digital Twin:
dynamic evolution and real-time interaction. In comparison to traditional models,
these characteristics lend distinct features to the credibility evaluation process of
Digital Twins.

4.1 Multi-dimensional Evaluation

Equipment is characterized by multi-disciplinary, multi-level, and long life cycle.

The Digital Twin of equipment is a complex model that involves multiple disciplines,
hierarchies, and model integrations. The credibility evaluation of Digital Twins needs
to be carried out from six perspectives, as shown in Fig. 2: model life cycle, model
form, model transparency, multi-disciplinary domain, model granularity, and model
scale.

(1) Model life cycle: A Digital Twin is developed around mapping the whole life
cycle of equipment (design, processing, testing, operation, and maintenance).
Meanwhile, the Digital Twin model also has a complete life cycle for any life
cycle stage of equipment, including design, construction, use, evolution, and
management. Each stage has an independent model form and function focus.
Therefore, it is necessary to investigate the life cycle stage of the twin model
and select the appropriate essential credibility metrics and evaluation methods
according to the user requirements and model characteristics.
(2) Model form: Different life stages correspond to various model forms, such as
conceptual models, mathematical models, and simulation models. The concep-
tual model is a formal interpretation of requirements. The mathematical models
are divided into mechanism models and non-mechanism models. Simulation
models are the code expression of conceptual models and mathematical models,
A Framework for the Credibility Evaluation of Digital Twins 75

Fig. 2 Multi-dimensional perspectives of equipment models

so the inspection of its logic and standardization is an important content of

credibility evaluation.
(3) Model transparency: According to understanding of a model mechanism, there
are three types of transparency: black box, white box, and gray box. The mecha-
nism of the white-box model is completely clear. The black-box model fits only
the input and output data without using the mechanism. The gray-box model has
an unclear mechanism, mainly presented by empirical formulas. Its evaluation
should consider both the mechanism and the output.
(4) Multi-disciplinary domain: Equipment models may include single or multiple
domains like electromechanical, hydraulic, pneumatic, thermal, and control,
which are presented in the form of continuous models and discrete models to
describe geometric, physical, behavioral, and cognitive models. For models in
different disciplines, specific theories and tools should be applied to evaluate
the relevant ability of the model according to the domain knowledge.
(5) Model granularity: A DT model of an equipment usually has a multi-granularity
structure including parts, components, equipment, subsystems, and systems.
Each granularity model type has different evaluation emphases according to its
unique characteristics. Compared with the parts or components, systems have
strong emergence and complex functions. The evaluation of systems should be
closely combined with the needs and be conducted hierarchically.
(6) Model scale: The scale of equipment includes temporal scale and spatial scale.
The scale measures the scope of the investigation, while the granularity measures
76 H. Lu et al.

the size of basic constituent units within the scope. There is a strong corre-
lation between spatial scale and granularity. Due to the limitation of evalua-
tion resources, the smaller the investigation granularity, the smaller the corre-
sponding spatial scale, and the larger the spatial scale, the larger the investigation
granularity.
In addition to the above six dimensions, the complexity of the model structure,
the model composition, the relationship between models, the model interoperability,
and the interaction between models and real-time data should be considered in the
evaluation. Therefore, it is necessary to have a set of index system involving multi-
dimensional elements and corresponding hierarchical evaluation methods.

4.2 Coupling of Modeling–Application–Evaluation

The real-time evolutionary nature of Digital Twins couples the construction process
and application process of twin models. Unlike the traditional basic model where
there is a clear boundary between construction and application, the twin model keeps
changing along with the changes of physical equipment during its applications. Thus,
the evaluation of the construction and application should be regarded as part of the
whole process, which can be called as a “modeling-application-evaluation cycle”.
As shown in Fig. 3, the key output and state data of the running equipment are
collected in real time. According to the difference between the real and virtual data,
the synchronized twin model identifies the biased components and conducts online
evolutionary calibration. Thus, the consistency between the model and the equip-
ment in terms of mechanism, structure, and parameters could be maintained. Once
the credibility of the twin model meets the requirements, various simulation-based
applications are allowed to be executed on the equipment.
The effect of the dynamic evolution process is shown in the credible evolution
box in Fig. 3, where the blue line refers to the states and outputs of the physical
equipment and the red line shows those of the twin model. The physical equipment
is influenced by the environment and teams including the equipment suppliers, the DT
users, and the evaluators. Compared with the equipment, the initial twin model still
has difference which will increase with the equipment running. The trigger module
calculates the incoming difference in real time and triggers the evolution process
according to the preset threshold. Then, the evolution module receives the messages
and starts assimilating the data.
As the number of loop iterations increases, the output error of the twin model will
be gradually reduced and eventually converged with the equipment output. In this
process, the evolution method will be optimally adjusted according to the obtained
credibility. Optimization aspects include the precision, over-advancement, and light
weighting, so that the twin model can continuously achieve higher credibility under
the given resource conditions.
A Framework for the Credibility Evaluation of Digital Twins 77

Fig. 3 Coupling of modeling–application–evaluation in DTs

From the above analysis, DT is a combination of modeling–application–evalua-

tion processes. The credibility evaluation of the DT is an online evaluation of the
dynamic model. For a given evaluation time window, the DT has one credibility
value. As the evaluation time window slides on the real-time axis, there is a corre-
sponding series of changing credibility values. The trend of the credibility value
is more indicative of the actual credibility than the value under any specific time
window.

4.3 Integrated Evaluation of Data and Models

The real-time interaction between the equipment and the twin model is the premise
of DT dynamic evolution, which includes forward calibration and feedback opti-
mization. Forward calibration transfers data from the equipment to the Digital Twin,
including real-time collection of equipment operating information, data generated
from various software, manually updated data like text drawings. Feedback optimiza-
tion delivers optimized control signals and scheduling planning from the Digital Twin
to the equipment, to promote the iterative updating of the equipment. As the equip-
ment is in operation, the user’s application requirements for the equipment may
change from time to time, which will deeply affect the credibility connotation of
the relevant Digital Twin. So, it is necessary to recognize the changes in the user’s
requirements and update the method of credibility evaluation in time.
78 H. Lu et al.

The two directions of real-time interactions are relatively independent but closely
related. Forward calibration is the foundation of credibility insurance. The quality and
timeliness of the data in this process will deeply affect the credibility of subsequent
predictions and simulations. Feedback optimization is the purpose of DT. If the
performance of the system is degraded in this step, it is likely to lead to a vicious cycle.
So actual feedback should be implemented cautiously with sufficient simulation
demonstration and risk prevention measures. During the evaluation, the independence
and relevance of the two directions should be fully considered from the perspectives
of both data and the twin model.

5 Overall Framework for Credibility Evaluation

As shown in Fig. 4, the credibility evaluation of the Digital Twin is mainly composed
of three parts: the construction of the index system, the basic evaluation, and the
evolution evaluation. The Digital Twin is a simulation model constructed around the
requirements. So, the multi-dimensional and multi-level index system required for
the evaluation should be obtained by analyzing the user requirements for the model.
The appropriate evaluation methods are then constructed to form a methodological
framework, by combining the characteristics of the model and investigation results
of the requirements.
On a given evaluation time window, the twin model is composed of a sequence
of basic models arranged on the time axis. The model evolution which occurs at a
time point generates a basic model for a new time point according to the previous
basic model; thus, an update of the twin model is completed. The basic evaluation
involves every basic model in the model sequence. The first basic evaluation needs
to be conducted in a comprehensive and detailed manner. The subsequent evalua-
tions are more lightweight and could only focus on the changed components. The

Fig. 4 Framework of DT credibility evaluation

A Framework for the Credibility Evaluation of Digital Twins 79

evolution evaluation examines the changes in the basic model sequence. The correct-
ness of each basic model and the timeliness of the overall basic model sequence are
influenced by each other. In general, the pursuit of higher correctness would lead to
lower timeliness. The examination of uncertainty gives the confidence distribution
for each quantitative value. The overall combination of correctness, timeliness, and
confidence distribution forms the credibility of the twin model.
When a closed loop is formed by connecting the basic model with the equipment,
the evaluator should continuously conduct credibility evaluation throughout the life
cycle of the model according to the basic characteristics of each stage. The following
section will focus on the framework of the index system.

5.1 The Index System

Although simulation technology has been developed for more than 70 years, there
is no unified index system about credibility.
Due to the complexity of the object, the credibility of the twin model needs to be
examined from multiple aspects. Synthesizing the various types of index systems on
credibility in different fields, and considering the characteristics of Digital Twins,
this chapter proposes a framework of credibility evaluation index system for Digital
Twins (Fig. 5). The superscript of the index symbols is the classification to which it
belongs, and the subscripts is the meaning of the indexes. The system consists of four
major parts, which can be further refined and expanded according to the meanings
of the indexes below, combined with the needs and characteristics of the evaluation
object.

(1) Data Indexes ID

Data is the foundation of Digital Twins construction and has a significant impact on
DT credibility. DT data is often characterized by real-time, high concurrency, and
multi-source heterogeneity. Data evaluation examines both the source that generates
the data and the data itself. The indexes are shown in Table 1.

Fig. 5 Framework of DT credibility evaluation indicator system

80 H. Lu et al.

Table 1 Data indexes

Object Index Meaning
Data Compatibility Generate, collect, and transmit data that meets performance
source IcD requirements
D
Authority Iao Equipment manufacturers, personnel organizations, and software
are authoritative or reliable
Data D The data arrives at the system in time and is consistent with the
Authenticity Iae
body reality
Sufficiency IsD The data fully contains the required information and is
standardized and easy to access

(2) Static Indexes IS

Evolution of a twin model is the generation of a new model by making parametric,

structural, or mechanistic changes to a basic model. Both the basic models before
and after evolution can be considered static. Static indexes are for the consistency
of the virtual-real mapping of the basic model and are important for subsequent
dynamic index. The static index is shown in Table 2 which considers the following
areas: analysis and correlation of requirements and models, single-dimensional key
attributes, and cross-dimensional coupling effects.

Table 2 Static indexes

Object Index Meaning
Requirement-model mapping Normalization InS QUOTE ZJgf Using standardized formal
methods to characterize
requirements and models.
Completeness IcS QUOTE ZJwz The requirements are
systematically and completely
correlated and mapped to a
multi-dimensional model.
Unidimensional attributes S QUOTE ZJ
Accuracy Iua All unidimensional
zq1
mechanisms and parameters in
Fig. 2 accurately reflect the
equipment attributes.
Independence IiS QUOTE ZJdl Each unidimensional model, as
well as its internal attributes
and functions, is relatively
independent.
Multi-dimensional coupling S QUOTE ZJ
Accuracy Ima The description of coupling
zq2
effects is concise and
consistent with actual results.
Boundedness IbS QUOTE ZJyj The coupling process has
reasonable time or value
boundary constraints.
A Framework for the Credibility Evaluation of Digital Twins 81

(3) Dynamic Indexes IDy

When the basic model has formed a closed loop with the physical equipment, the
model parameters, structure, and mechanism will be updated based on the data
collected from the physical system. Model evolution is an important feature which
separates Digital Twin models from traditional models. As shown in Table 3, the
dynamic indexes are for the evolution process, which mainly examine whether the
virtual model correctly follows the physical object, whether the response to various
evolutionary events is timely, and whether the whole evolution trend converges.
Based on the multi-dimensional correlation analysis of the model in the preparation
of Sect. 5.2.1, the dynamic changes in the model structure and mechanism can be
parameterized for consistency evaluation.

(4) Supporting indexes ISu

The above three indexes are primary, while the supporting indexes are often used as
references. These include maturity of capabilities and maturity of models [43], which
are more difficult to obtain and has indirect impact on model credibility (Table 4).

Table 3 Dynamic indexes

Object Index Meaning
Evolution Discriminative Be able to correctly identify the changed part of the model when
trigger Dy the physical system undergoes changes
Id
Dy
Sensitivity Ise The minimum data fluctuation required to achieve identification
of changing parts
Dy
Evolution Timeliness It The ratio of evolution complexity to the time spent on
process corresponding evolution
Convergence The convergence of differences between model and equipment
Dy under the influence of evolution and uncertainty
Icv
Dy
Evolution Systemic Isy Evolution participation level of parts that are related to the
results trigger one and affect the requirements
Dy
Correctness Icr The degree of difference between the multi-time window output
of the measurable object and the actual output

Table 4 Supporting indexes

Object Meaning
Capability maturity Including team collaboration, management standardization, team authority,
etc.
Model maturity Including reusability, security, robustness, sustainability, etc.
82 H. Lu et al.

Fig. 6 Framework of DT credibility evaluation methods

Based on the quantitative mapping relationship between the requirements and

the model, the degree to which each index value is within the acceptable range of
the requirements is the degree of correctness. For the index values involving any
processes, the degree to which the time consumption is within the acceptable time
range is the degree of timeliness.

5.2 Methodology Framework for Credibility Evaluation

The methodology framework for credibility evaluation of Digital Twins is shown

in Fig. 6. The evaluation process is mainly composed of four segments: prepara-
tion, method construction, basic evaluation, and evolution evaluation. The core idea
and focus of each segment of the evaluation work are described below. At the end
of the evaluation, a dynamic multi-dimensional distribution of credibility values
considering correctness, timeliness, and uncertainty will be formed. The integration
of values is based on the demand-model mapping. The obtained credibility values
could provide an important reference for further optimization of the construction and
decision-making application of Digital Twins.

5.2.1 Preparation

(1) Analyze the model composition of the evaluation object. According to Fig. 2, a
multi-dimensional decomposition diagram of the complex model (Fig. 7) should
A Framework for the Credibility Evaluation of Digital Twins 83

Fig. 7 Multi-dimensional
decomposition diagram of
complex models

be constructed from six perspectives. Also the cross-dimensional associations

and the associations among the same dimensions should be described by the
characterization methods, such as directed graphs. Based on this description, the
geometric structure variation index or integrated analysis methods, such as AHP
and Bayesian network, can be used to parametrically characterize the changes
of the structures and mechanisms.
(2) Analyze the requirements for the evaluation object. In this process, methods like
formal analysis and mathematical modeling are applied. The fuzzy requirements
described in natural language are transformed into semi-formalized descrip-
tions, further specified qualitatively or quantitatively. Finally, each requirement
is fully decomposed and mapped to the six-dimensional perspectives.

5.2.2 Methodology Construction

(1) Determine the specific index system. The specific indexes are further refined by
considering both the characteristics of the model and the requirement analysis. A
complete multi-dimensional hierarchical index system would be formed based
on the index system shown in Fig. 5. Furthermore, the acceptable value range
of each index is determined under the guidance of quantitative requirement
analysis, which would serve as a benchmark for the correctness of the subsequent
evaluation.
(2) Select appropriate methods. Eighty-one existing credibility evaluation methods
are sorted out and organized into in eight categories (Table 5). Different indexes
can be quantified and assessed using the appropriate methods.
For each single index, it is necessary to select the appropriate method or combi-
nation of methods from the list by considering the life cycle stage where the model
serves, model characteristics, and user requirements. There is still a lack of evalu-
ation methods for some indexes, which is also an important part of future study. In
84 H. Lu et al.

Table 5 List of credibility evaluation methods

Top level Secondary level Top level Secondary level
Software evaluation F1 Code inspection F11 Results comparison Sensitivity analysis
F5 F51
Algorithm analysis F12 Probability statistics
F52
Execution test F13 Deterministic data
analysis F53
Dataset validation F14 Evaluation of error
effects F54
Subjective methods F2 Manual review F21 Integrated analysis Analytic hierarchy
F6 process F61
Turing Test F22 Bayesian network
F62
Expert scoring F23 Gray correlation
analysis F63
Empirical comparison Complex network
F24 analysis F64
Semi-formal methods Consistency check F31 Evolution analysis F7 Timeliness analysis
F3 F71
Graphical analysis F32 Accuracy analysis
F72
Formal method F4 Inference by induction Smoothness
F41 analysis F73
Assertion check F42 Systematic analysis
F74
Theory of evidence F43 Data examination F8 Data certification
F81
Formal review F44 Data validation F82
Logical inference F45 Data source analysis
F83

addition, appropriate integration methods need to be chosen according to primarily

three types of situations: multi-disciplinary coupling, multi-dimensional integration,
and multi-temporal fusion.

5.2.3 Basic Evaluation

(1) Evaluation of Static Models

1. Requirement-model mapping evaluation. According to the analysis in
Sect. 5.2.1, examine whether the user requirements are correctly analyzed
or normatively characterized for the normalization index InS . Afterward,
A Framework for the Credibility Evaluation of Digital Twins 85

examine whether the system architecture and functional design of the

model sufficiently support the requirements for the completeness index IcS .
Subjective methods such as F21 and F23 are mostly used.
2. Unidimensional evaluation. The static model is initially divided into specific
evaluation objects of single life cycle and single granularity under the overall
spatial and temporal scales. The model is then gradually decoupled into unit
models from three perspectives: model forms, disciplinary domains, and
transparency of mechanisms. Using a combination of F1, F3, F4, and F5
methods, the indexes under each single dimension are assessed for accuracy
index IuaS
and independence index ISi . This process is the focus of the basic
evaluation, which requires multi-disciplinary expertise and analytical tools.
3. Coupling evaluation: F13, F32, F5, and other methods are mainly used to
examine the coupling effect between multiple disciplinaries and multiple
dimensions. The accuracy index Ima S
QUOTE ZJzq2 of the coupling effect
characterization can be examined from the unity of the conserved quantity,
the coverage of the associated factors, and the consistency of the feedback
effect. The boundedness index QUOTE ZJyj IbS of the coupling can be exam-
ined from the consistency of the boundary conditions and the convergence
of the coupling.
It is also necessary to integrate the results of multi-dimensional and multi-
disciplinary single-index evaluations into a coherent hierarchical network
of credibility values according to the intrinsic correlations of the system. In
this case, multiple integration analysis methods in F6 are used.
(2) A credibility evaluation of the multi-source heterogeneous data (Fig. 8) is
performed to ensure that the basement for subsequent model evolution is correct.
This step considers both the data source and data body. The examined data source
includes collection devices, human-related organizations, and software applica-
tions. The main purpose is to examine the performance of the related equipment,
the rationality of the layout, the interference resistance, and the data standard-
ization, etc. The key is to examine the compatibility index IcD with methods
D
such as F81, F83, and to take the authority index Iao of the data source as a
reference. Data body refers to the data that arrives at the system. The key is
D
to evaluate its authenticity index Iae through data comparison, cross-validation,
and to examine its sufficiency index IsD in combination with traceability, value
consistency. Methods such as F5 and F82 are mainly used.

(3) Evaluation of supporting elements. For the maturity of capabilities and models,
a comprehensive examination of the process of their evaluations and improve-
ments under multiple application scenarios and different needs is performed.
These experiences could serve as references for credibility estimation of unprac-
ticed scenarios under the same scope, which can be referred to in the literature
[43].
86 H. Lu et al.

Fig. 8 Multi-source heterogeneous data of DT

5.2.4 Evolution Evaluation

Evolution evaluation focuses on the evolution process of the simulation model. There
are some similarities between the dynamic characteristic analysis of the traditional
basic model and the evolution evaluation of the twin model. However, they are signif-
icantly different. On the one hand, the evolution of the model makes the dynamic
characteristic curves change continuously. So the evolution evaluation cannot only
analyze the curves at a specific time point, but rather, it has to comprehensively
consider the entire sequence of curves on a period of time window. The key element
to be examined during evolution evaluation is the change process of the basic model
sequence. On the other hand, the twin model shows the property of modeling–appli-
cation–evaluation cycle as mentioned in Sect. 4.2. Thus, during an evolution process,
the object keeps being assessed and the twin model may change accordingly. Then,
the evaluation criteria should be changed as well.
In general, there are three types of credibility evaluation triggering methods corre-
sponding to the changes of the object: time-driven, event-driven, and hybrid-driven.
There are also four types of time windows, which are shown in Fig. 9. In the scenario
of online dynamic evaluation of Digital Twins, it is necessary to divide the time
windows according to the actual triggering mechanism, periodicity, and overlapping
Dy
situation. Generally, the evolution-driven discriminative index Id is evaluated using
Dy
F23, F31, and the sensitivity index Ise is analyzed using methods such as F51. For a
Dy
single time window evolution process, its timeliness index It can be checked using
Dy
methods such as F71, and its convergence index Icv can be analyzed using methods
Dy
such as F73. For the corresponding evolution result, the systemic index Isy of the
Dy
evolution can be assessed using methods such as F74, and the correctness index Icv
of the evolutionary result can be analyzed using methods such as F72.
The dynamic integration method is based on multiple time windows. The intrinsic
correlation between multiple triggering mechanisms is analyzed. Then, the dispersion
of the weights on the multiple time windows are obtained, so that the dynamic index
values could be continuously synthesized from the basic index values.
A Framework for the Credibility Evaluation of Digital Twins 87

Fig. 9 Multiple time windows’ based credibility evaluation

Uncertainty is examined at the last step. For each index value, the data, evalua-
tion methods, and management methods involved in the acquisition process should
be examined. Both the inherent uncertainty, including systematic parameter errors,
environmental perturbations, and the cognitive uncertainty, including inappropriate
model forms and empirical parameter errors are included. The failure probability of
the assessed value brought about by the uncertainty is then quantified and uniformly
mapped to the distribution of confidence intervals. By combining this distribution
with its degree of correctness and timeliness, the degree to which the model actu-
ally meets the demand, which is the degree of credibility, could be obtained. Based
on the requirement-model mapping, the single-index value, multi-dimensional inte-
gration value, and dynamic evaluation value are transformed into the final dynamic
multi-dimensional distribution of the credibility.

6 Example of Credibility Evaluation of a Digital Twin

The process of implementing the credibility evaluation methodology framework is

illustrated based on the Digital Twin of the ROKAE industrial robotic arm constructed
by the research team. The experimental environment is shown in Fig. 10. The hard-
ware used is the XB4 robotic arm, the Mech-mind Nano depth camera, the Dahuan
88 H. Lu et al.

Fig. 10 DT environment of a robotic arm

AG95 claw, sensors such as thermometers, hall sensors, voltage and current sensors,
and communication devices such as routers and edge gateways. The software used
is Unity, ANSYS Workbench, C#.
The credibility evaluation started with a multi-dimensional decomposition of the
robotic arm DT model by the modeling experts. The process is followed by analyzing
the requirements for “grasping a target and placing it in a specified area”. The two
key inspection points are then identified: “stable grasping with appropriate force”
and “timely routing and obstacle avoidance”. Afterward, the model is decomposed
step by step from qualitative to quantitative to obtain the specific index system and
the acceptable range of requirements. In this step, the normalization index InS and
completeness index IcS of the requirements-model mapping are assessed.
A unidimensional evaluation focusing on multi-disciplinary domains and simu-
lation forms was then conducted. Among the multi-disciplinary domains, geometry,
kinematics, mechanics, control, electricity, and magnetism have a strong influence
on the grasping requirements, so the correctness of their disciplinary models is the
focus of the evaluation. Meanwhile, the correctness of the system structure, code and
interface specification, and logical correctness of the simulation models are the main
investigations of the polymorphic dimension.
Thereafter, the compatibility index IcD of the acquisition communication device
D
and the authority index Iao of data sources such as vendors and built-in software
of the robotic arm were evaluated using formal review and expert scoring methods.
D
The authenticity index Iae of historical modeling data, real-time status data, and
requirement data was evaluated in depth using a variety of data validation methods.
A Framework for the Credibility Evaluation of Digital Twins 89

The overall technology maturity was assessed using subjective and formal methods
in conjunction with historical data related to model development and application.
The uncertainty of each evaluation component was analyzed using dataset validation
and result comparison methods.
After completing the basic evaluation, a multi-disciplinary and multi-dimensional
integration was performed using a modified DT theory of evidence. The integration
was based on the following Dempster synthesis rule:

0,

(A = )
m(A) = Ai ∩Bj =A m1 (Ai )m2 (Bj ) ,
1−K
, (A = )
  
where K is the conflict factor which meets K = Ai ∩Bj =A m1 (Ai )m2 Bj < 1 and
reflects the degree of conflict between the various pieces of evidence.
On this basis, the two key evolution processes of the robotic arm attitude change
and real-time deformation of the gripped object are assessed. The process and results
of the calibration of each parameter and the geometrical structure change are system-
atically examined. The time-frequency domain analysis and the convergence anal-
ysis are comprehensively used to obtain the six index values of the whole evolution
process. The quantitative value of the model credibility is given by combining with
the uncertainty analysis.
In the credibility synthesis process, the weights of each index are given mainly
based on the analysis of scene requirements and experts’ experience. The weights
assigned to the data indexes are relatively small because the experimental environ-
ment is relatively simple and stable, and the robotic arm as well as the related acqui-
sition equipment has high reliability. High weights are assigned to the indexes related
to unidimensional and coupled evaluation due to the high impact of the correctness
of the discipline model on the grasping requirements. Because the real-time calibra-
tion of the robotic arm attitude changes and the deformation of the grasped object
has a greater impact on the grasping and obstacle avoidance needs, high weights are
also assigned to the evolution-related indexes. The other indexes have less influence
on the application requirements of the robotic arm twin model and are uniformly
assigned smaller weights. A multi-objective optimization algorithm will be used in
the subsequent study to determine the optimal weight allocation for each index.
To validate the credibility evaluation methods, four types of DT models with
different credibility are constructed to form a basic model benchmark, as shown in
Fig. 11. The grasping effects of the four types of DT models are shown in Fig. 12.
The horizontal axis is the time axis, and the vertical axis is the distance between
the grasped object and the end point. The evaluation calculation results give the
order of model credibility as Model 3 > Model 4 > Model 2 > Model 1, which is
consistent with the experimental performance results. Thus, the effectiveness of the
credibility evaluation framework proposed in this chapter is preliminarily validated.
It is necessary to expand the library of benchmark models in terms of equipment type,
model mechanism, and complexity of scenario requirements, so that the credibility
evaluation framework could be developed and validated in a wider scope.
90 H. Lu et al.

Fig. 11 Grasping experiments on four types of models with different credibility

Fig. 12 Experiment results of four types of models with different credibility

7 Conclusions

This chapter analyzed three typical characteristics of Digital Twins: dynamic evolu-
tion, real-time interaction, and system complexity. On this basis, the definition of
credibility of Digital Twin is given. Then, a set of credibility evaluation index system
is proposed based on previous researches and practical experiences. The general
process of DT credibility evaluation is described afterward. The credibility evaluation
framework is validated through experiments.
A Framework for the Credibility Evaluation of Digital Twins 91

In the future research, the evaluation index system will be further extended and
refined so that credibility can be described more comprehensively and accurately.
Moreover, some single-index evaluation methods will be improved or supplemented,
such as uncertainty quantification method and credibility integration method. In
addition, more effective quantification methods will be developed for the credibility
evaluation of the evolution model. The refinement and combination of evaluation
methods would be further explored according to the commonality and difference of
the geometric appearance model, the system structure model, and the mechanism
model. The benchmark library of Digital Twin models would be extended to better
validate the generalization capability of the credibility evaluation framework. An
evaluation software system would be developed based on the framework to support
the automatic credibility evaluation to a certain extent.

References

1. Zhang L, Zhou L, Horn BKP (2021) Building a right digital twin with model engineering. J
Manuf Syst 59:151–164
2. Shafto M, Conroy M, Doyle R et al (2010) Technology area 11: modeling, simulation, infor-
mation technology and processing roadmap. NASA Office of Chief Technologist, November
2010
3. Boschert S, Rosen R (2016) Digital Twin-the simulation aspect. Springer International
Publishing, Cham, pp 59–74
4. Bielefeldt B, Hochhalter J, Hartl D (2015) Computationally efficient analysis of SMA sensory
particles embedded in complex aerostructures using a substructure approach. In: ASME 2015
conference on smart materials, adaptive structures and intelligent systems. ASME, New York,
USA, V001T02A007
5. Liu S, Bao J, Lu Y et al (2021) Digital Twin modeling method based on biomimicry for
machining aerospace components. J Manuf Syst 58(Part B):180–195
6. Zhang L (2020) Cold thinking about the digital twin and the modeling and simulation techniques
behind it. J Syst Simul 32(4):1–10
7. Zhang L, Lu H (2021) Discussing digital twin from of modeling and simulation. J Syst Simul
33(5):995–1007
8. Tuegel E J (2012) The airframe digital twin: some challenges to realization. In: 53rd AIAA/
ASME/ASCE/AHS/ASC structures, structural dynamics and materials conference. 20th AIAA/
ASME/AHS adaptive structures conference 14th AIAA, 1812
9. Huang J, Gheorghe A, Handley H et al (2020) Towards digital engineering: the advent of digital
systems engineering. Int J Syst Syst Eng 10(3):234–261
10. Zhang L, Wang K, Laili Y et al (2022) Modeling & simulation based system of systems
engineering. J Syst Simul 34(2):179–190
11. Rabeau R (2013) Credibility in modeling and simulation. In: Simulation and modeling of
systems of systems, pp 99–157
12. Department of Defense (2007) DoD Modeling and Simulation (M&S) Management. Directive
Number 5000.59, 8 Aug 2007
13. Sargent RG (2015) Model verification and validation. In: Modeling and simulation in the
systems engineering life cycle: core concepts and accompanying lectures. Springer London,
pp 57–65
14. Grieves M (2019) Virtually intelligent product systems: digital and physical twins. In: Flumer-
felt S et al (eds) Complex systems engineering: theory and practice. American Institute of
Aeronautics and Astronautics, pp 175–200
92 H. Lu et al.

15. Lim KYH, Zheng P, Chen CH (2020) A state-of-the-art survey of Digital Twin: techniques,
engineering product lifecycle management and business innovation perspectives. J Intell Manuf
31(6):1313–1337
16. Liu M, Fang S, Dong H et al (2021) Review of digital twin about concepts, technologies, and
industrial applications. J Manuf Syst 58:346–361
17. Lu Y, Liu C, Kevin I et al (2020) Digital Twin-driven smart manufacturing: connotation,
reference model, applications and research issues. Robot Comput Integr Manuf 61:101837
18. Sun Y, Luo H, Das SK (2012) A trust-based framework for fault-tolerant data aggregation in
wireless multimedia sensor networks. IEEE Trans Dependable Secure Comput 9(6):785–797
19. Jiang J, Han G, Wang F et al (2014) An efficient distributed trust model for wireless sensor
networks. IEEE Trans Parallel Distrib Syst 26(5):1228–1237
20. Miao CL (2014) Research on trustworthiness problems in wireless sensor networks. University
of Science and Technology of China
21. Mao HM (2016) Research of ultrasonic sensor detection model based on credibility evaluation.
Shenyang Jianzhu University
22. Liao M, Liu J, Meng Z et al (2021) A sins/sar/gps fusion positioning system based on sensor
credibility evaluations. Remote Sens 13(21):4463
23. Levelt P, Caramazza A (2007) The Oxford handbook of psycholinguistics. Oxford University
Press, USA
24. Barber A, Robert JS (2010) Concise encyclopedia of philosophy of language and linguistics.
Elsevier
25. Tang JB (2009) Research on the credibility of warfare simulation system. Graduate School of
National University of Defense Technology
26. Liao Y, Feng XJ, Zhang J et al (2003) Research on calibration, verification and validation
(VVA) technology for strategic missile system modeling and simulation. In: Proceedings of
the 2003 National Academic conference on system simulation, pp 187–192
27. Yang M, Zhang B, Wang ZC et al (1999) VV&A and credibility evaluation of distributed
interactive simulation system. J Harbin Inst Technol 05:48–51
28. Kim CJ, Lee D, Hur S et al (2017) M&S VV&A for aerospace system development: M&S life-
cycle model and development paradigm. J Korean Soc Aeronautical Space Sci 45(6):508–516
29. Zhou JH, Xue JJ, Li HY et al (2020) Thinking on digital twin for weapon system. J Syst Simul
32(04):539–552
30. Zhang SL, Ye MC (2006) Simulation credibility evaluation method for missile weapon system.
Comput Simul 23:48–51
31. Zhen BT, Hu Y (2015) Research on foreign weapon and equipment modeling and simulation
credibility assessment methods. Qual Reliab 2:6–9
32. Sim SH, Lee JC (2014) On improving the verification, validation and accreditation process by
including safety requirements in M&S-based development of weapon systems. J Korea Safety
Manage Sci 16(4):123–131
33. Beydoun G, Low G, Bogg P (2013) Suitability assessment framework of agent-based software
architectures. Inf Softw Technol 55(4):673–689
34. Acar E (2015) Effect of error metrics on optimum weight factor selection for ensemble of
metamodels. Expert Syst Appl 42(5):2703–2709
35. Li N, Dong L, Zhao L et al (2021) A credibility evaluation method for complex simulation
systems based on interactive network analysis. Simul Model Pract Theory 110:102289
36. Ferson S, Oberkampf WL (2009) Validation of imprecise probability models. Int J Reliab Saf
3(1–3):3–22
37. Li W, Chen W, Jiang Z et al (2014) New validation metrics for models with multiple correlated
responses. Reliab Eng Syst Saf 127:1–11
38. Dornheim H, Brazauskas V (2011) Robust–efficient credibility models with heavy-tailed
claims: a mixed linear models perspective. Insurance Math Econ 48(1):72–84
39. Liang J, Bai Y, Bi C et al (2013) Adaptive routing based on Bayesian network and fuzzy
decision algorithm in delay-tolerant network. In: 2013 IEEE 10th international conference on
high performance computing and communications & 2013 IEEE international conference on
embedded and ubiquitous computing. IEEE, pp 690–697
A Framework for the Credibility Evaluation of Digital Twins 93

40. Hu XL (2011) Dynamic data driven simulation. SCS M&S Mag 5:16–22
41. Wang K, Zhang L, Jia Z et al (2023) A framework and method for equipment digital twin
dynamic evolution based on IExATCN. J Intell Manuf 1–13
42. Wang K, Zhang L, Cheng H et al (2023) A lifelong learning method based on event-triggered
online frozen-EWC transformer encoder for equipment digital twin dynamic evolution. IEEE
Internet Things J
43. Zhang L, Liu Y, Laili Y et al (2020) Model maturity towards modeling and simulation:
concepts, index system framework and evaluation method. Int J Modeling Simul Sci Comput
11(03):2040001
On the Importance of Simulation
and Digital Twin for Industrial
Applications

Stefan Boschert, Christoph Heinrich, Vincent Malik, Roland Rosen,

and Uliana Soellner

Abstract The term Digital Twin has become ubiquitous recently as virtual repre-
sentation of a physical asset to leverage business opportunities, i.e., with a dedicated
purpose. It consists of very heterogeneous data and can take very different shapes,
from the collection of datasheet information together with a visual reproduction of
the real object via the compilation of operational data to the application of physics-
based simulations or predictions based on artificial intelligence (AI). Obviously, any
combination thereof is also possible. The applications of Digital Twins are spread
over the whole lifecycle of products and systems including the interaction of Digital
Twins at different hierarchical levels. In addition, Digital Twins can create a connec-
tion between different value chains. An example is the exchange of Digital Twins
along the supply chain with a link to the shop floor, which then enables the execution
of cooperative simulations for different purposes. In our book chapter, we investigate
the role of simulation for a Digital Twin with respect to value chains, the lifecycle
phases, and their business value. Furthermore, in the age of the Industrial Metaverse
the importance of simulation for Digital Twins is getting even more important, as
immersion, interaction, and collaboration in real-time play a greater role, making
it necessary to have fast and good enough predictions of the behavior of industrial
assets. Novel algorithms and machine learning (ML) methods enhance the capabili-
ties of the traditionally engineering-focused simulation tools. The mentioned aspects
will be illustrated with several examples of real-life industrial applications.

Keywords Digital Twin · Simulation · Industrial Metaverse

1 Introduction

The efficient development and optimal operation of automated systems are well-
known, but still challenging. Automated systems are mostly complex technical
systems that are controlled by (automation) software and increasingly networked with

S. Boschert (B) · C. Heinrich · V. Malik · R. Rosen · U. Soellner

Technology, Siemens AG, 85748 Garching, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 95

M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_5
96 S. Boschert et al.

one another, which transforms technical systems into cyber-physical systems (CPS).
Concrete examples are production systems, their sub-systems, and components in
the process and manufacturing industry.
In addition to the consideration of quality, time, and cost targets, the flexibility
of these systems in operation is increasingly coming to the fore. At the same time,
the variety of product types is increasing and the batch sizes are becoming smaller.
Short-term changes in orders as well as changes in supplies are necessarily to be
taken into account. Instead of supplementing classic automation, Digital Twins are
increasingly being named and realized to fulfill these functions. In this chapter,
we consider Digital Twin applications that have simulation-based and data-driven
function modules and can improve operation in various ways (cf. [1, 2]).
The term Digital Twin has grown historically, introduced in the context of PLM as
to collect the information about a product across its lifecycle [3] and first mentioned as
Digital Twin in a NASA Roadmap in 2010 [4]. Nowadays, it contains all information
about an asset for the entire life cycle and becomes even a synonym for the asset
administration shell in the Industrie 4.0 context [5]. Unfortunately, no consistent
definition is used in practice. It is mostly shaped by the perspective and use of the
respective user.
In the following, we discuss different lifecycle aspects of Digital Twins and moti-
vate the concept of bundling best practices for simulation-based Digital Twins in the
context of engineering and automation. Then we discuss several novel applications
of the Digital Twin and give concrete examples of their realization. We conclude
with a discussion on the future development of Digital Twins.

2 Digital Twin Aspects and Goals

In recent years, we have seen an increasing digitization and the triumph of Internet
technologies. Most of the products and systems are created in a digitally supported
development process. Standardized interfaces and data formats are the foundation
of integrated toolchains for the creation and execution of models during all life-
cycle phases [6]. In operation phases, Internet technologies for data acquisition, data
transformation, and data analysis (e.g., by artificial intelligence (AI) methods) have
become cheap, easy to install and use, and have been commoditized. The pervasion
of virtual and augmented reality (VR/AR) into daily-life applications, called Meta-
verse, is currently creating a demand for interactive and intrusive solutions. All this
leads to an increased interest and awareness of Digital Twins and their purpose and
goal.
On the Importance of Simulation and Digital Twin for Industrial … 97

2.1 Digital Twin in the Lifecycle

Digital representations are found everywhere during the lifecycle of a system or

product. Already during the initial conceptual phase, first ideas, mainly in the form
of requirements, are collected mostly in a semi-formal way. In a more systematic
way, the requirement specification is created, to describe the basic behavior of the
final system – a first comprehensive description. During the following design and
engineering steps this description is continuously modified and complemented with
new information and data. First (system) architectures, CAD representations of sub-
components, and simulation models to virtually test the components against the
given requirements are developed, see Fig. 1. Often different versions and variants
of each of those digital artifacts must be considered as well. Even if the real system
or product is not yet built, this set of information gives a reasonable picture of the
final product and its behavior—at least under certain pre-considered situations. They
form a virtual representation of the knowledge about the physical asset—an initial
Digital Twin [7, 8].
The Digital Twin does not only contain descriptive information like CAD models,
but also executable models in the form of simulation models, based on physical laws
or heuristics like previous experience. This allows us to calculate the behavior of
the physical asset for a pre-defined situation. It must be noted that these models are
specific for their intended use, and it must always be checked if the model is suitable
(i.e., it satisfies its prerequisites) for the current situation under consideration. If this
is not the case, a new model or a modification of an existing similar (simulation)
model must be created to answer the specific design question or to validate specific
properties. This newly created model becomes then part of the Digital Twin as well.
As such, the Digital Twin becomes more and more comprehensive.
Another aspect to consider is that most systems are not completely developed
in all details by a single stakeholder. We must consider the supply chain as well.
Partners and suppliers are involved who supply components or sub-systems. These
suppliers in turn have suppliers and both use Digital Twins to describe their system.
This results in a linked chain and an exchange of digital artifacts respectively Digital
Twins. This finally forms a Digital Twin ecosystem.
Looking at the manufacturing of a product, Digital Twins can support it as well.
On the one hand, the compatibility of the product with the production line can be
tested in advance. This can be done for example by checking different geometrical or

Fig. 1 Digital Twin along the life cycle

98 S. Boschert et al.

functional parameters of the product with the capabilities of the production system.
On the other hand, the Digital Twin of the production line can be used to test out
different operating/manufacturing strategies. There is also a certain analogy in the
process industry. Here, too, the behavior of the process plant can be tested and
validated.
The production, however, is a stage in the lifecycle of the product where the
characteristics of the Digital Twin change significantly. Before the production of the
physical asset, only the “theoretical”/desired behavior of the product in a general-
ized environment is considered assuming estimated operating conditions. As soon
as the product is realized, individual aspects of each of the products are added to
its Digital Twin. This could be the manufacturing history like “manufactured on
machine x with quality y”, “used parts/substances from supplier z” or “batch id
number and environment conditions like temperature or humidity”. Especially in
the case of mass products, each of the products carries along its own production
and operational history, even if the basic product description is initially the same.
In [9], this transformation is called instantiation of the Digital Twin: The Digital
Twin prototype (DTP) that was developed during the design becomes a Digital Twin
instance (DTI) of each individual product. Consequently, the content of the initial
DTP is also available as a common set of information in all instances. This common
set of knowledge is complemented by individual aspects of each instance, like the
individual and specific environment conditions or operational modes. Using these
individual data, individualized predictions for each asset can be made. This requires
on the other side that the data for each asset must be treated individually, which can
be done locally—on the edge—or in a more centralized way or in a mixture of both
extremes, depending on the availability of computing resources.
Based on this set of information/data, executable models can be created to describe
a specific behavior aspect of the asset. Usually, these models are either numerical
simulation models based on a mathematical description of the relevant physics, or
data-based models that are trained with available sensor data. The desired accuracy
of these models widely depends on the necessities of the specific use case. As these
requirements are often opposite to each other, one has to compromise, e.g., execution
speed versus accuracy, resulting in a “good enough” simulation model for the specific
purpose. Therefore, it could also happen that models of the same physical effect are
part of the Digital Twin in different model fidelity.
Further challenges arise as the model must cope with real sensor data. Measure-
ments include uncertainties that should be considered in the simulation model as
well. And more importantly, the simulation must process the sensor information in
(near) real time to give fast enough results.
The Digital Twin will lead to novel software applications in operation and service
phases and, for example, through the establishment of edge and cloud architectures
to new business models. During operation, more and more data, not only from local
sensor measurements but also operating data from connected IT systems, are avail-
able and can be considered in these applications. Technologically, this means that
simulation models are also combined with data-based methods and AI techniques.
On the Importance of Simulation and Digital Twin for Industrial … 99

Typical applications of such simulation-based Digital Twin applications support

systems for operation, where for example operation alternatives can be tested in
advance on the virtual system so that the operator can choose the best option. In
some situations, even a fully automatic optimization of the operation is possible,
if the models and data provide a sufficient accuracy and execution speed. Another
typical application could be the support of maintenance actions. Running a simulation
in parallel to the real asset offers the possibility either to detect derivations from the
expected (i.e., simulation) and the real behavior (obtained from sensor readings) or to
improve operational aspects. Concrete applications are virtual sensors, calculations
of remaining useful lifetime, failure detection, and identification or improved spare
part management.
One fascinating idea with Digital Twins is that different types of digital infor-
mation like descriptive and executable models can be used in a linked way. For a
machining tool, a service application can be built which calculates the optimal process
time and process parameters (=executable model), depending on the state of the
machine, tool, and orders (=descriptive data). Another example is an online produc-
tion planning application, which optimizes the production steps of each product
unit in a line based on the actual order times, current machine states, and related
maintenance plans.
Looking at the supply chain perspective, the linked digital artifacts along all
tiers form a root network of data. This makes it possible to trace this information
and answer questions such as “Which parts does my system consist of?”, “Which
suppliers manufactured it?”, and “Of what quality is it?”, and “Which footprint does
it have?”. When considering these topics, often the term “Digital Thread” is used,
which emphasizes the temporal and logical connection of the information. Digital
Threads help to navigate through them in order to solve specific challenges like
product improvements swiftly and reliably.

2.2 Bundling Best Practices for a Novel Simulation-Based

DT Approach

From a research point of view and to derive technical ideas for the preliminary devel-
opment, we propose a novel simulation-based Digital Twin concept which extends
traditional simulation approaches and improves most likely the usage of Digital
Twins. This vision of the future has developed further in recent years (see [5, 10,
11]) and manifests itself in the following characterization:
The (in previous papers called next generation) Digital Twin refers to a description
of a component, product, system, infrastructure, or process by a set of well-aligned,
descriptive, and executable models. It is a semantically linked collection of all rele-
vant digital artifacts, including design and engineering data, operational data, and
behavioral descriptions.
100 S. Boschert et al.

The Digital Twin has several characteristic features:

. It evolves along the lifecycle—use and content should be pre-planned but is open
for additional, un-planned artifacts.
. It integrates the currently available and commonly required information and
knowledge.
. It addresses (proto)type and instance aspects.
. It is synchronized with the physical world—the type of synchronization changes
over the different lifecycle phases.
. It offers an efficient way to generate new services and applications—besides
extended solutions in development new offerings in operation and service promise
new business opportunities.
With respect to the focus of this paper on simulation-based Digital Twin applica-
tion, we see new or at least increased business value in a realization of this approach
even if several individual aspects need to be considered. This will be illustrated in
the following section.

3 Novel Simulation-Based Digital Twin Applications

3.1 Digital Twins in Networked Value Chains

The development and use of technical systems almost always take place in a network
of development partners, suppliers of parts, components, and products, and of
customers who use these systems or in turn integrate them into their systems and sell
them to their customers. At the end of this chain is the end customer who uses or
operates this system. An example illustrates these connections of partners, e.g., the
component could be an electric drive designed for a control valve. This product is
manufactured in the manufacturing plant and installed in a production plant. Or the
valve is used by system manufacturer to manufacture a complete process module. It is
an open, dynamically changing network. After all, information is already exchanged
between partners or suppliers and customers today (and in the past). The Digital Twin
approach will significantly increase the effectiveness and efficiency of information
exchange, as more complex knowledge can be exchanged using models. The artifacts
exchanged can be Digital Twins or, in terms of the description given above (”a set
of well-aligned, descriptive, and executable models”), a subset of digital artifacts.
Regarding the simulation part, the previously mentioned modularization is essential
here, as it improves reuse and thus increases cost-effectiveness. To put it briefly, the
previous fragmented exchange will be replaced by comprehensive puzzle pieces in
the future. In implementation, their unambiguous description of their content and
their interfaces is important to be able to use them as independent simulation-based
Digital Twins or to connect them with other twins (puzzle pieces). In this way, these
Digital Twins are transformed into independent products that are subject to specific
On the Importance of Simulation and Digital Twin for Industrial … 101

Fig. 2 Crossing of value chains in production, adapted from [14]

quality requirements [12, 13]. Encapsulation in modules also improves maintain-

ability, because if an update like the change of a physical asset is necessary, the
corresponding module can be updated or replaced in a targeted manner.
One application focus for Digital Twins is the creation, operation, and modern-
ization of production facilities. Figure 2 (adapted from [14]) shows the value chains
of the product, the production system, and the supply chain that meets on the shop
floor. Each value chain represents related views and tasks [15]:
. Product Lifecycle Management of the product (PLM): This is the systematic
process of managing the product over its whole lifecycle. This includes ideation,
design, engineering tasks as well as its production and service up to disposal of the
product. This product is the output of the production system under consideration. A
second type of products is the means of production (PLM of equipment). These are
machines, actors, sensors, automation systems, and so on which are the elements
of the production system itself.
. Production System Lifecycle Management (PSLM): It handles all processes which
are relevant for the production system and its context like the factory building or
plant area. The PSLM contains the phases design, engineering, construction and
commissioning, operation (including service and maintenance), and end-of-life
aspects.
. Supply Chain Management (SCM): SCM is the management of the flow of goods
and services. An important aspect for the production is the logistics of raw material
and supply parts as well as warehouse aspects and delivering of finished products.

To achieve economically successful and sustainable production, the combination

of information from these value chains is required. The Digital Twin is an important
success factor here. It contains various information and thus transports knowledge
across the lifecycle phases and, analogous to the representation in Fig. 2, it enables
a linking of the information, e.g., between product development and production.
102 S. Boschert et al.

A digital convergence of product and production system expands the opportunities

in the planning and development of production systems as well as in their (virtual)
commissioning and operation. Potential benefits of (simulation-based) DT applica-
tions lie in the construction, modification (e.g., product-specific reconfiguration),
and expansion of production. This includes matching product requirements and the
capabilities of production resources (e.g., machines and process units) together with
the electrical and automation equipment properties of the plant in planning and vali-
dation tasks. In case of short-term production disruptions or changes in the supply
chain (e.g., changed deliveries in terms of time, quantity or quality as well as changed
orders), alternatives can be determined and evaluated online. This requires a media
discontinuity-free transfer of all Digital Twin artifacts which are required to imple-
ment simulation-based applications. To achieve this goal, practical implementations
for the above-mentioned modularization concept must be achieved. An increased use
of semantic models will help to link all information.
Increasing the economic benefit from the further use of measurement and oper-
ating data is summarized by the buzzword data-to-value. The simulation-based
Digital Twin will expand the (so far) often purely data- and AI-based methods by
including simulation models from the development phases of the production equip-
ment and the production system. This extends the performance capabilities of diag-
nostic and service applications and allows online controls for operation optimization
[10, 11]. The synchronization of the Digital Twin with its physical asset is decisive
for the technical feasibility. The physical asset’s status must be identified adequately
and in a purpose-oriented way by the input data coming from the real asset. The
technical implementation needs to consider the specific required update cycles and
transfer rates.
Such a Digital Twin, combining information from different domains, will continue
to reinforce the development and operation of cyber-physical systems, which are
increasingly moving toward autonomous systems. These aspects are presented in the
next section.

3.2 Digital Twin Contributions Toward System Integration

Autonomous systems have become popular for several years, especially through
assistance systems in vehicles. In addition to the application in the automotive
industry, this development direction is also relevant for other complex technical
systems, e.g., in manufacturing and process industries.
According to Bekey’s definition [16] “Autonomy refers to systems capable of oper-
ating in the real-world environment without any form of external control for extended
periods of time”, autonomous systems are intelligent machines (or systems again) that
execute high-level tasks without detailed programming and without human control.
They know their capabilities and skills, and their internal state, and they use sensors
to perceive their environment and the current situation. They can make decisions
based on scenario analysis on how to proceed. In order to make this happen, the
On the Importance of Simulation and Digital Twin for Industrial … 103

autonomous systems will need access to models that are as fine-grained as the use
case requires and their own behavior in interaction with their environment in the
real world—these are typical tasks of a Digital Twin [2]. This once again proves the
relevance of the topic.
Full autonomous behavior of systems will remain a vision for years, but we will
continuously approach this goal. The concept of Digital Twins will be very useful
for this. An intermediate step is the realization of cyber-physical systems (CPS)
which communicate with their environment and can fulfill more complex tasks. The
economic realization of such CPS will require new development methods, and the
Digital Twin approach will support it.
A typical approach for system development is to start from requirements and define
functional blocks to fulfill them. These functions are further sub-divided into logical
systems which are finally realized in their physical representation. Mechatronic
systems become more and more cyber-physical, which means that the communication
aspect either within the system but also between different systems becomes more and
more important. Therefore, the system under consideration becomes larger and finally
harder and harder to understand in all details. The quest for the optimal system design
becomes challenging. One approach could be to split the full system into several sub-
systems, e.g., functional blocks and optimize their operation. Especially in classical
mechatronic systems like automotive industry this approach delivers good results,
e.g., by sub-dividing the system “car” into many different components which could
be finally produced by a large ecosystem of suppliers, each one specializing on some
components and optimizing the components by themselves. Only by integration of
the components to the full system are the interactions (physical or communication)
between the separate components realized in a pre-defined way—mostly determined
by the initial functional and logical sub-division of the full system.
However, as the communication aspect becomes more dominant in cyber-physical
systems, it also becomes more challenging to identify components that are mutually
related and could benefit from each other. This becomes obvious at the example of
a logistics system, where on the one side of a conveyor belt the belt is loaded from
different sources, and on the other side, the material must be sorted into different
buckets. If on the one side the sources are running at maximum speed (are optimized
for maximum performance), the sink side could run into trouble as several buckets
get overloaded—even if the sink side runs on maximum as well as some buckets
are full and must be emptied. In this situation, it would be helpful to slow down the
performance of the source side—or modify the mixture in such a way that the sinks
are more evenly accessed—to maximize the overall performance of the system. Here
source and sink form a symbiotic pair by coordinating their behavior to the benefit
of both components and finally the overall system.
A central part for the realization of such symbiotic effects is to establish a commu-
nication link between the components. This, however, requires that the relevant
components are already identified. Approaches to identify such symbiotic pairs can
be found, e.g., in [17]. The identification of the symbiotic pairs should have an impact
on the design of the control system as well, as a direct communication is established
more easily if the components are controlled by the same controller. Obviously, an
104 S. Boschert et al.

initial partition (probably dominated by geometric reasons) of the system must be

adjusted after the identification of symbiotic interdependencies.
From the benefit point of view, the operation of systems that can solve complex
tasks independently is attractive. A common approach for production (equally in
the manufacturing and process industries) is that production systems become more
modular. These modular units will act more independently. To do this, they need to
exchange information with each other. This is one of the basic principles of Indus-
trie 4.0. With the definition of the Asset Administration Shell (AAS) [6], vendor-
independent exchange of information is being driven forward by corresponding stan-
dardizations. This information relates, among other things, to machines and their
components as well as other information such as order data. Since Digital Twins are
required for decision making in autonomous systems, the AAS must also provide
services for using the Digital Twins. One aspect of the simulation of sub-models is
that it makes intensive use of co-simulation concepts [18].
The development of CPS and the use of Digital Twins are an economically rele-
vant and a technologically interesting intermediate step. In production, the aim is
to create cyber-physical production systems that are structured more modular and
enable better autonomous operation of the production units. Their benefit is the
increasing flexibility. Changes in the production process and intralogistics can be
implemented ad hoc. Like cars, the technical possibilities and the trust in the deci-
sions of these systems will grow step by step. Digital plant companions which leave
the final decision to humans represent an intermediate step here [19].

3.3 Simulation in the Metaverse Era

Only recently has the Industrial Metaverse (IMV) as evolution of the Digital Twin
era emerged. Whereas the term Metaverse is already used in the gaming market, it is
developing in three additional fields. In the consumer market, it represents an avatar-
centric virtual world for socializing and entertainment, whereas in the enterprise
market, it is used in connection with real-time collaboration to facilitate office tasks
employing large language models in the future. In contrast to that, the IMV as the
third field aims at solving real-world problems, increasing productivity, shortening
time-to-market and, thus, adding value to industry. Prognosis of market segmentation
even sees the IMV at around 100 billion US$, outcompeting the other two fields by
a factor of more than 2–3 [20].
Concretely, the IMV is a space to experience Digital Twins of industrial assets with
the key features immersion, collaboration, and interaction where the combination
goes beyond Digital Twins. The visualization of the Digital Twin in its context and
realistic environment leads to new insights. Meetings between different stakeholders
to jointly review Digital Twins facilitate problem solving where changes can be
directly applied. Moreover, interactive simulation of the Digital Twin behavior and
On the Importance of Simulation and Digital Twin for Industrial … 105

faster-than-real-time prediction leverage predictive maintenance to a new level. The

connection to the physical assets to monitor and analyze in a closed loop is a major
differentiator of the IMV.
We see the IMV as a convergence of various key technologies to create indus-
trial virtual worlds, e.g., Internet of Things, edge and cloud computing, artificial
intelligence machine learning, user experience, networks (5G/6G), and blockchain.
But the key enabler of the IMV is the photorealistic, physics-based, and real-time
capable Digital Twin where in turn simulation is the main ingredient, see Fig. 3.
Simulation based on the conservation of physical laws, like, e.g., the finite element
method (FEM) for structure mechanics, is physically correct by construction and
thus makes it possible to get the accuracy needed for the respective application. This
drawback of AI-based methods is currently a strong research field with the goal to
fulfill physical laws also for, e.g., neural networks.
The IMV creates added value along the lifecycle and connects the different phases.
In design and engineering, it helps to perform interactive simulations with collabora-
tive review and evaluation. Photorealistic environments for synthetic data generation
speed up virtual testing and validation tremendously. In production, new assets and
control software can be tested in the virtual worlds and deployed to the real hardware
after successful virtual commissioning. Remote control of assets and their service
reduce downtimes in operation, saves costs, and overcomes the shortage of skilled
workers by intuitive visualization of data and interaction with software and hardware
as well as remote training.
Although the IMV rather represents a vision where foundations are still being laid,
first implementations already show its great potential. By using simulation and Digital
Twin to plan a new factory in Nanjing the productivity could be increased up to 20%

Fig. 3 Digital Twin of a production line based on a simulation model and its photorealistic
visualization [21]
106 S. Boschert et al.

compared to conventional and comparable factories [20]. This, in turn, has impact
on space efficiency, required material, and energy consumption. Consequently, the
IMV can be also seen as a key enabler for sustainability.
Another example will be explained in the example section.

4 Digital Twin Realizations

4.1 A Digital Twin for Design and Operation of a Complex

Chemical Reactor

The progress in science and technology leads to improvements in manufacturing

processes. One of the modern methods of manufacturing is additive manufacturing
(AM). AM allows to manufacture complex functional structures. This in turn makes
it possible to adapt geometries with respect to physical processes and realize new
advanced applications with high performance. Examples of such applications could
be cooling systems or chemical reactors, like the methanol synthesis reactor described
in the following. Conventional reactor geometries for methanol synthesis are quite
simple, but they have a significant size and are mainly used in industrial chemical
plants. However, methanol synthesis can be applied to store electrical energy from
renewable energy sources, so-called Power-to-X reactors [22]. In this case, the reac-
tors work under transient conditions, are significantly smaller, and therefore require
additional studies for a novel reactor design. AM is a promising method which
allows compact design of reactors thus saving material and as a result improving the
sustainability of the power-to-chemical process. The challenges of reactor design for
methanol synthesis are operation requirements and boundary conditions: high pres-
sure, temperature control, condensation of chemical reactions products. The Digital
Twin allows to understand the physical processes in the reactor in more detail and to
find the optimal design.
Figure 4 shows the development chain of optimal design. At the beginning, an
initial design concept is developed, which is based on the analysis of existing designs
and a comprehensive understanding of physical processes. Once the geometry for
the first prototype is ready, its characteristics can be studied by physical test or
simulations. In many cases, the numerical modeling can estimate the properties of
component prototype with less time and material efforts than physical tests. This is
especially important at early stages of design development. And even if simulations
are not able to describe all multi-physics processes without simplifications, they
allow us to estimate the geometry-functionality dependency. In a chemical reactor
for the methanol synthesis, we have the following processes: chemical reactions on
the catalyst surface, phase transition of reaction products methanol and water, and
temperature conditioning. Since the system is under high pressure, the stiffness of
the structure must be considered during developing of the reactor design. All these
processes can be described as multi-scale and multi-physics processes. Including all
On the Importance of Simulation and Digital Twin for Industrial … 107

Fig. 4 Development circuit for the functional design

physical and chemical processes in detail in a single simulation model will lead to
high complexity and computational costs. Therefore, it is reasonable to use simpli-
fication methods like homogenized models. To ensure that the selected models are
applicable and correct for describing a specific process, it is necessary to validate
these models with physical tests or high-resolution multi-physics simulations (if
high-fidelity simulations are available for the respective process). For the model
verification, a physical test or high-resolution simulations can be performed for a
representative volume or simplified geometry. The validated models are then used to
simulate and optimize the complex structures and geometries of functional compo-
nents. As such, already during the early design phase a set of specific simulation
models is created that form the initial Digital Twin of the reactor.
Let us look at the workflow from detailed representations to homogenized models
for 3D simulations of methanol synthesis reactor. The methanol synthesis is a
complex chemical process that can be described by means of kinetic models [23]
received from fitting the measurement results. In the high-resolution multi-physical
simulations, the geometry of the catalyst is reproduced in detail. Therefore, the
properties such as pressure drop or effective thermal conductivity of the catalyst
bed for a homogeneous model are calculated. Similarly, the basic properties of the
porous metallic structures in the reactor are defined. Using the high-resolution simu-
lations and tests the porous materials were analyzed, taking into account the geometry
including all pores and collected information. Then this information is used to find
out the parameters for the homogenized models like the Darcy-Forchheimer model
for pressure drop [24, 25]. The difficulty so far lies in the calculation of the dynamic
capillary pressure with detailed pore geometry, since this simulation is transient,
requires a high computational effort, and depends on the contact angle, which does
108 S. Boschert et al.

not always remain constant during the solid–liquid interaction. Currently, the high-
resolution multi-physics simulations are not available for the porous structures in
our chemical reactor. As a first approximation, the Leverett model [26] was used
to describe the capillary behavior in a porous structure. For the description of the
turbulent flow in the reactor, Reynolds-averaged Navier–Stokes equations (RANS
equations) with k-ε turbulence model [27, 28] are used, which provide results of
sufficient accuracy for given flow boundary conditions in the reactor. To describe the
phase transition a model with Eulerian averaging of the transport equations for the
additional phase interaction was used [29, 30]. This model solves the conservation
equations and assumes a common pressure field for the phases. It is important to
mention that the main goal of the models is to describe the process quantitatively
and not the micro-scale process reproduction (Fig. 5).
The simulation models that use homogenization of small-scale phenomena and
parametrization of geometry provide results during reasonable time for the industry
and allow to build an automatic circuit “CAD-simulation-optimization-CAD” to get
an optimal novel design of functional structures.
The created 3D multi-physics simulations for the understanding of detailed
processes and geometry optimization can be also applied to determine the missing
input data and boundary conditions like pressure drop in the catalyst bed and heat
transfer coefficients for one-dimensional simulations (1D simulations). The 1D simu-
lations allow not only to calculate the chemical transformation in a reactor under given
conditions but also can consider the reactor as a part of a chemical plant and define
the operation points for an entire facility. The local 3D simulation of the reactor
can be included in an overall system model to optimize the operation of the facility.
On the other hand, the system model provides more accurate boundary conditions

Fig. 5 Example of the connection between high-resolution simulations and modeling for methanol
synthesis reactors
On the Importance of Simulation and Digital Twin for Industrial … 109

for the local model. Therefore, the example of the methanol synthesis shows that
the development of design for a functional component is an iterative process using
Digital Twin models of varying degrees of scale. These models improve during this
iterative process and are then the base for a proper operation of the whole plant.

4.2 Health Monitoring for Drive Train Components

For a health monitoring application, a specific aspect of Digital Twin is utilized that
focuses primarily on modeling and simulating the behavior of a physical system or
process (Behavior Twin). This can be used to predict how the system will behave in
different scenarios and to support decisions or optimizations. This approach is applied
at different stages of the development to ensure the functionality and reliability
of complex systems, particularly in operation and service for health monitoring of
critical components.
The knowledge of the health status of electrical drive train components, such as
converter and motor, is important for safe operation and for planning the service
activities. To calculate the aging, knowledge of the temperature of critical compo-
nents (converters electronic, bearing, stator insulation) during the operation is essen-
tial. Using numerical simulation models, the calculation of health and remaining
useful life by processing of operational data and estimation of temperatures of crit-
ical components becomes possible. It eliminates the need for direct measurements,
which are often unfeasible due to limited access to the components and the huge
volume of operational and sensor data generated.
The workflow for monitoring the health of the drive train is shown in Fig. 6. This
process consists of several steps which contribute to the development and deployment
of a comprehensive health monitoring service. The first step involves the creation of
detailed models for each critical component of the drive train: (a) the application, (b)
the motor, and (c) the converter. These models describe components’ behavior and
consist of various digital artifacts, including system architectures, CAD representa-
tions of sub-components, materials, and control models. In the next phase (system
integration), the component models are combined to a model of the entire drive train
system. This step allows to simulate the interactions between the different compo-
nents, evaluate the system’s performance, and calculate the temperature of critical
components. To create a robust health monitoring service, the aging models are
integrated into the system model. These models consider the wear that drive train
components undergo over time and allow to estimate the components’ health and
remaining useful life.
It is important to know that these models are developed for a specific application
and need to be precisely evaluated to ensure that they are applicable for another
scenario. If a model does not meet the defined requirements, it is necessary to create
a new model or modify an existing model.
The execution of the developed service usually takes place on edge computers
(d) close to the data source. The most important advantages are low transmission
110 S. Boschert et al.

Service development
a) b) c) d)
e) Aging model

Model integration
Development of component models

Application Motor Converter

Fig. 6 Workflow for the health monitoring of the drive train

latency, reliability, and security. The other option is to run it as a cloud service. The
main reasons for this are scalability, flexibility in updates and maintenance.
In summary, the Digital Twin technology provides an innovative and physics-
based solution for the calculation of stress, temperature, and damage levels in compo-
nents without direct measurements. By applying the correlation between component
use, characteristics, and operating conditions, these predictive maintenance models
provide the way for safer and more efficient operations as well as the development
of more robust components.

4.3 Manufacturing-X and Catena-X

Manufacturing-X is a German initiative, started by business, politics, and academia

to digitalize supply chains in the industry [31] and to realize an open and global
data ecosystem which allows companies to share data across production and supply
chains. The general goal is to implement a federated, decentralized, and collaborative
data ecosystem for smart manufacturing. Concrete aims are
. “the reorganization of value networks so that they can react quickly to incidents
(resilience),
. new business models, a circular economy, and increased efficiency (sustain-
ability), and
. digital innovation that will secure and expand the global leadership of German
industry (competitiveness)” (see [32]).
The initiative pursues a cross-sectoral solution approach (see Fig. 7) and drives the
internationalization by implementing global standards and setting up international
communities, e.g., the International Manufacturing-X Council.
A blueprint for Manufacturing-X is the Catena-X flagship project, which is
publicly funded by the Federal Ministry for Economic Affairs and Climate Action
(BMWK) and was started in August 2021 with a duration of 3 years [33]. Goal of
On the Importance of Simulation and Digital Twin for Industrial … 111

Fig. 7 Cross-sectoral approach to realize a networked industry [31]

the project is to create a collaborative and open data ecosystem for the automotive
industry. Digital Twins are a core element to realize the provision and exchange of
data [34].
The information required for the exchange between partners in the data ecosystem
is described as semantic data models. For transmission, they are filled with the specific
data to be exchanged, i.e., instantiated. This principle is implemented via asset admin-
istration shells [35]. In Industrial Digital Twin Association (IDTA) terminology, these
data models correspond to sub-models. These sub-models are supplemented by infor-
mation about the object to which this data relates. Both parts together form the asset
administration shell. In the understanding of IDTA, the asset administration shell and
Digital Twin are synonyms [6]. Digital Twins are therefore the means of transport
for software solutions in data ecosystems. They are used, for example, to exchange
information along the supply chain, as the next example illustrates.

4.4 Online Control and Simulation

Online Control and Simulation (OSim) is the name of a use case in the German public-
funded project Catena-X. The goal of the research development is a software solution
which allows a distributed, collaborative material flow simulation of production and
logistics. The strategic goal is the increased resilience of supply chains by faster,
more precise reactions in case of short-term disruptions [36]. OSim is one of five use
cases in the Catena-X Business Domain Resiliency [37].
112 S. Boschert et al.

Supply chains are subject to various disruptive factors, and it is important to react
flexibly and make effective adjustments. An obstacle to this is the often insufficient
exchange of information between the individual partners in the chain. Even if indi-
vidual logistics or production partners carry out their planning locally optimally,
without overarching coordination this can lead to inefficient supply chains.
The OSim solution approach aims to simulate the behavior of the entire supply
chain, including the effects of short-term disruptions or changes, at the material flow
modeling level [38]. The simulation is not carried out in a holistic, monolithic model,
but as a distributed simulation in which each partner simulates their processes and
sequences. The exchange of information with suppliers and customers necessary for
the respective simulation takes place via standardized data models. It maintains the
data sovereignty of the partners and guarantees compliance with antitrust regulations.
Through collaborative simulation, it becomes possible to react to changes in
advance and improve efficiency across the entire supply chain by the increased
observation horizon.
All partners involved in the supply chain simulate their respective processes and
actions (production or logistical sequences) based on current information from the
shop floor and its IT systems (e.g., manufacturing execution systems). The results of
the simulation, essentially delivery data, and quantities of parts produced or trans-
ported, are passed on to the immediate successor in the supply chain, who in turn
carries out a simulation of their material flow processes based on this possibly
changed information. In OSim, a “horizontal” exchange of information across the
supply chain is combined with a “vertical” exchange of the material flow simulation
of its production or its logistics processes (e.g., transport times). Through iterative
simulations by all partners and information exchange upstream and downstream,
improved efficiency of the entire supply chain is achieved in a collaborative way.
The core of this approach is an OSim manager, which is operated by every partner
involved. The OSim manager is a software application that records all the information
that is necessary for the local execution of a material flow simulation and supplies it to
the locally executed simulation. After the simulation has been carried out, the OSim
manager provides the results of the simulation to the partners again. For the secure
and data-sovereign transmission of information, the Eclipse Data Space Connector
(EDC) is used in accordance with the Catena-X regulations [39].
The executed OSim manager with its connected local simulations forms a Digital
Twin of the material flow of the entire supply chain. Also, every exchange of infor-
mation between the partners uses (in the IDTA understanding) a Digital Twin, which
essentially contains the results of the simulation.

4.5 The Industrial Metaverse for Battery Production

Use cases for the Industrial Metaverse are numerous and manifold. Although it
still represents a vision, Siemens and Freyr have already showcased in [40] how
On the Importance of Simulation and Digital Twin for Industrial … 113

Fig. 8 Digital Twin of a battery plant including production and assembly lines [40]

manufacturers can work in the future based on the example of a battery production
site, see Fig. 8.
By creating the Digital Twin including an Industrial Metaverse experience of
a factory that hasn’t been built yet, confidence can be built among stakeholders
including investors. Being immersed into the future production site, seeing the
production lines virtually commissioned, and having transparency on the perfor-
mance of the whole plant creates trust in the capability of the corresponding company
and reduces the risk of wrong investments significantly.
Moreover, the IMV provides the means to not only display KPIs from the whole
plant down to the component level but also to optimize them. Here, simulation plays a
crucial role. For example, in the critical mixing process (see Fig. 9), the mixing degree
can be simulated by means of CFD (computational fluid dynamics) and particle simu-
lation to finally optimize the process time impacting the overall performance of the
plant. If this is not enough, then even the machine itself or machine components can
be optimized. Simulation and design optimization enable dedicated improvements
like the geometry of the stirrer. Thus, the IMV allows for virtually testing different
parameters before the plant is built or without compromising the actual production
of an existing plant. This becomes especially important in the operational phase of
a plant when anomalies are detected, and a root-cause analysis can be performed in
the IMV. Here, different stakeholders from different locations can virtually meet in
the plant, analyze operational data, and test corrective actions before applying them
in the real plant. With this approach, the costly downtimes of a plant can be reduced
to a minimum.
There are many more use cases where the IMV creates an added value in a factory.
In the assembly (see Fig. 9 on right) health and safety issues can be evaluated such as
the interplay between robots and human workers or test the ergonomics by simulation.
As energy consumption and the corresponding CO2 equivalent became a major KPI
in production, the IMV of a plant can also be used to create transparency and to
optimize scheduling to additionally save costs by making use of changing energy
prices.
114 S. Boschert et al.

Fig. 9 Simulated mixing process allows for optimization of mixing time and mixing degree [20]

5 Commenting on the Future Development of the Digital

Twin Approach

The term Digital Twin is now being used for more than ten years. Its roots lie in
simulation, other influences came and continue to come from areas such as data
analytics, the Internet of Things (IoT), and Industry 4.0. The latter claims to equate
the terms Digital Twin and asset administration shell [6].
A comparison between the definition given here by the authors and the Asset
Administration Shell shows great similarities. Information and data from different
sources are used and integrated. Knowledge capturing takes place; data and
simulation-based models play a central role here. Both approaches address the entire
lifecycle of assets, from design and development to operation and maintenance.
Furthermore, there is a clear relationship to the counterpart of the Digital Twin—
the real twin. These similarities justify using both terms synonymously and further
convergence will increase as further sub-models of the AAS are defined. To evaluate
the future development of the Digital Twin approach, it makes sense to take a focused
look at the core properties of both terms and the approaches behind them. The Asset
Administration Shell is a standardized framework that provides a digital represen-
tation of a physical asset. It serves as a comprehensive information model, which
includes in particular administrative and business-related information on the physical
asset. The framework defines a common language designed for seamless integra-
tion between different assets and interoperable communication between systems and
devices. The Digital Twin approach emphasizes the creation process of the digital
representation. This representation includes relatively detailed and dynamic behav-
ioral models consisting of detailed analysis, optimization, and automated support of
users in their different tasks in all life cycle phases. A Digital Twin refers not only
On the Importance of Simulation and Digital Twin for Industrial … 115

to physical assets, but also to planned real objects (from small components to large
infrastructures), technical and business processes and people.
The term Digital Twin is associated with two aspects that must be viewed in a
differentiated manner. On the one hand, this is the representation of the real coun-
terpart itself and, on the other hand, the use and application of the Digital Twin. The
first aspect includes elements of a scientific discipline that pursues technical improve-
ment of the models used and works on further developing the methods required for
this. The second aspect addresses the creation of software applications that directly
provide the Digital Twin as a concrete solution for a task. The types of solutions
range from simple tools to fully automated IT solutions.
This leads to a future vision of the Digital Twin that is characterized by several
elements. The importance of the models used will continue to increase. These must
be easier to combine in their usage and therefore more modular in their methodolog-
ical basis. The models must meet typical product characteristics. These are explicit
descriptions of the model contents and model limitations as well as assurances about
model quality in terms of maturity and stability. In order to ensure economic use,
these models must be easier to develop and maintain. This refers on the one hand to
the further use of the models or the Digital Twin for variants of the planned or real
counterpart and on the other hand to the successor generation of the real twin. This in
turn requires better integration into development tools and existing or new software
solutions for operation and service, which is also a task in the Digital Twin tech-
nology area. These more scientific and technical future aspects of the Digital Twin
continue approaches that were started in the 1980s with the CAx methods and are
now finding a future with the Industrial Metaverse. On the application perspective of
Digital Twins, another important element is the efficient generation of mostly soft-
ware applications which allow the direct usage of a Digital Twin for concrete tasks.
The Asset Administration Shell framework plays a crucial role in implementing
these goals. It enables the management of Digital Twins and their integration into
various applications. However, other objectives require improvements on technical
and scientific elements. On the Digital Twin application perspective, it is expected that
specific application solutions besides the AAS will be continuously developed which
are independent and directly marketable products. The Digital Twin and its further
development offer numerous opportunities and potential for companies in various
industries. By using models that have modularity and clear product descriptions as
well as integrating appropriate tools and applications, new uses can be opened up
and efficiency increased. The Digital Twin is therefore an important technology for
the future of industry to realize better insights and decision making for the planned
or existing real counterpart.
116 S. Boschert et al.

References

1. Kuehn W (2018) Digital twins for decision making in complex production and logistic enter-
prises. Int J Des Nat Ecodynamics 13(3):260–271. https://doi.org/10.2495/DNE-V13-N3-
260-271
2. Rosen R, von Wichert G, Lo G, Bettenhausen KD (2015) About the importance of autonomy
and digital twins for the future of manufacturing. IFAC-Pap 48(3):567–572
3. Grieves M (2002) Completing the cycle: using PLM information in the sales and service
functions [slides]. SME Management Forum, Troy, MI
4. Piascik R, Vickers J, Lowry D et al (2010) Technology area 12: materials, structures, mechanical
systems, and manufacturing road map. NASA Office of Chief Technologist
5. Wagner C, Grothoff J, Epple U, Drath R, Malakuti S, Grüner S, Hoffmeister M, Zimmermann P
(2017) The role of the Industry 4.0 Asset Administration Shell and the Digital Twin during the
life cycle of a plant. In: Proceedings of the 2017 IEEE conference on Emerging Technologies
Factory Automation (ETFA 2017), Limassol, Cyprus
6. Rosen R et al (2021) Die Rolle der Simulation im Kontext des Digitalen Zwillings: Virtuelle
Inbetriebnahme-Modelle als Mittler zwischen Phasen und Anwendungen? atp magazin, Bd 63,
Nr 04, S. 82–89
7. Boschert S, Rosen R (2016) Digital Twin—the simulation aspect. In: Hehenberger P, Bradley
D (eds) Mechatronic futures. Springer International Publishing, Cham, pp 59–74
8. Drath R, Malakuti S, Grüner S, Grothoff J, Wagner C, Epple U, Hoffmeister M, Zimmer-
mann P (2017) Die Rolle der Industrie 4.0 „Verwaltungsschale“ und des „digitalen Zwillings“
im Lebenszyklus einer Anlage - Navigationshilfe, Begriffsbestimmung und Abgrenzung. In:
Tagungsband zur Automation 2017. Langfassung auf Tagungs-CD (12 Seiten), VDI-Verlag,
Baden-Baden. ISBN 978-3-18-092293-5
9. Grieves M, Vickers J (2017) Digital Twin: mitigating unpredictable, undesirable emergent
behavior in complex systems. In: Kahlen FJ et al (eds) Transdisciplinary perspectives on
complex systems. Springer, Heidelberg
10. Boschert S, Rosen R, Heinrich C (2018) Next generation digital twin. In: Horvath I et al (eds)
Proceedings of TMCE 2018, S 209–218
11. Rosen R, Fischer J, Boschert S (2019) Next generation digital twin: an ecosystem for
mechatronic systems? IFAC-Pap 52(15):265–270. https://doi.org/10.1016/j.ifacol.2019.11.685
12. Barth M, Kübler K, Heinzerling T, Rosen R, Jäkel J (2020) Eine systematische Bewertung der
Qualität von Simulationsmodellen: Identifikation und Clustering von Qualitätskriterien. atp
magazin (06–07), S 68–75
13. Barth M et al (2021) Eine systematische Bewertung der Qualität von Simulationsmodellen,
Teil 2: Bewertungsmetrik und Validierung, atp magazin, Bd 63, Nr 04, S 56–63
14. Kalhoff J, Borchers E, Rosen R (2021) „Diskussionspapier: Production System Life-
cycle Management im Wandel der Fahrzeugindustrie“, Plattform Industrie 4.0, avail-
able at https://www.plattform-i40.de/PI40/Redaktion/DE/Downloads/Publikation/Diskussio
nspapier_kopa35c.html. Last access 04 June 2021
15. Bröcker S et al (2021) Prozesssimulation – Fit für die Zukunft?, Positionspapier der
DECHEMA und VDI. Available at: https://dechema.de/Positionspapier_Prozesssimulation-
path-123212,124930.html. Last access 04 June 2021
16. Bekey GA (2005) Autonomous robots: from biological inspiration to implementation and
control. The MIT Press. ISBN 9780262025782
17. Zeman K et al (2020) Symbiotic mechatronics: an alternative view on complex systems. In:
Proceedings of TMCE 2020, 11–15 May 2020, Dublin
18. SG AAS Teilmodell Simulation, Plattform Industrie 4.0 (2021) „Interoperabilität hoch zwei
für Simulationsmodelle. Available at https://www.plattform-i40.de/PI40/Redaktion/DE/Dow
nloads/Publikation/Interoperabilit%C3%A4t-hoch-zwei.html. Last access 04 June 2021
19. Fischer J et al (2021) Assistenzsysteme der nächsten Generation für Produktionssysteme, VDI
conference Automation 2021, Paper accepted
On the Importance of Simulation and Digital Twin for Industrial … 117

20. MIT Technology Review Insights (2023) The emergent industrial metaverse. https://ass
ets.new.siemens.com/siemens/assets/api/uuid:260d1ead-250b-45fb-b32f-caa9924c1ddd/The-
Emergent-Industrial-Metaverse.pdf
21. https://www.siemens.com/global/en/company/digital-transformation/industrial-metaverse/
what-is-the-industrial-metaverse-and-why-should-i-care.html
22. Kalz KF, Kraehnert R, Dvoyashkin M, Dittmeyer R, Gläser R, Krewer U, Reuter K, Grun-
waldt J-D (2017) Future challenges in heterogeneous catalysis: understanding catalysts under
dynamic reaction conditions. ChemCatChem 9:17–29
23. Lacerda de Oliveira Campos B, Herrera Delgado K, Pitter S, Sauer J (2021) Development of
consistent kinetic models derived from a microkinetic model of the methanol synthesis. Ind
Eng Chem Res 15074
24. Darcy H (1856) Dalmont V (ed) Les Fontaines Publiques de la Ville de Dijon: Exposition
et Application des Principes à Suivre et des Formules à Employer dans les Questions de
Distribution d’Eau (Paris)
25. Forchheimer P (1901) Wasserbewegung durch Boden. Z Ver Dtsch Ing 45:1781–1788
26. Leverett MC (1941) Capillary behavior in porous solids. Trans AIME 142:151
27. Lien FS, Chen WL, Leschziner MA (1996) Low-Reynolds number eddy-viscosity modelling
based on non-linear stress-strain/vorticity relations. In: Proceedings 3rd symposium on
engineering turbulence modelling and measurements, 27–29 May, Crete, Greece
28. Teli SM, Mathpati CS (2020) Computational fluid dynamics of rectangular external loop airlift
reactor. Int J Chem Reactor Eng 18(5–6):20200009
29. Drew DA, Passman S (1998) Theory of multicomponent fluids. Springer, New York
30. Ishii M (1975) Thermo-fluid dynamic, theory of two-phase, eyrolles
31. Plattform Industrie 4.0. Manufacturing-X. Available at https://www.plattform-i40.de/IP/Nav
igation/EN/Manufacturing-X/Manufacturing-X.html. Last access 19 Oct 2023
32. Plattform Industrie 4.0. White Paper on Manufacturing-X. Available at https://www.plattform-
i40.de/IP/Redaktion/EN/Downloads/Publikation/Manufacturing-X_long.html. Last access 19
Oct 2023
33. Federal Ministry for Economic Affairs and Climate Action (BMWK). Projektsteckbriefe. Avail-
able at https://www.bmwk.de/Redaktion/DE/Downloads/P-R/projektsteckbriefe.pdf?__blob=
publicationFile&v=1. Last access 19 Oct 2023
34. Catena-X (203) Semantic layer/digital twins. https://catena-x.net/en/offers-standards/sem
antic-layer. Last access 19 Oct 2023
35. IDTA, Industrial Digital Twin Association, information available at https://industrialdigitalt
win.org/en/technology. Last access 19 Oct 2023
36. Eclipse Tractus-X. Osim Kit. Available at https://eclipse-tractusx.github.io/docs-kits/kits/
OSim%20Kit/Adoption%20View%20OSim%20Kit. Last access 19 Oct 2023
37. Eclipse Tractus-X. Business domain resiliency. Available at https://eclipse-tractusx.github.io/
docs-kits/next/Resiliency/. Last access 19 Oct 2023
38. Wolfrum P, Haager D, Isinger T, Rosen R, Kutzler T (2023) Kollaborative Materialflusssim-
ulationen zur Steuerung der Verknüpfung von Supply Chain und Shopfloor - ein Catena-X
Use-Case. In: Automation, Baden-Baden
39. Catena-X, EDC—The Central Component. Available at https://catena-x.net/en/offers/edc-the-
central-component-1. Last access 19 Oct 2023
40. Siemens FREYR Industrial Metaverse showcase. YouTube. https://www.youtube.com/watch?
v=LQyKc6SI5u8
A Formal Framework for Digital Twin
Modeling, Verification, and Validation

Mamadou Kaba Traore , Simon Gorecki , and Yves Ducq

Abstract In today’s technological landscape, Digital Twins have emerged as pivotal

technologies, offering unprecedented opportunities for innovation and optimization
across various industries. Despite numerous on-going research and development
initiatives, the verification and validation (V&V) of the Digital Twin (DT) remains
a major scientific obstacle. In order to disambiguate the DT concept (as various
interpretations exist), this chapter first formalizes it, using the Discrete Event System
Specification (DEVS) approach. Then, following a review of the state-of-the-art in
V&V, it examines three avenues for developing methodologies for V&V of DTs:
software engineering-based V&V, simulation-based V&V, and formal method-based
V&V. That way, this chapter provides a framework for DT understanding and a
roadmap for future research on DT V&V.

Keywords Digital Twins · Verification · Validation · DEVS · Formal

Framework · Smart Systems

1 Introduction to Digital Twin Concepts

The notion of “smart everything” is surfacing alongside the continuous digitalization

of society, industries, health, educational, and urbanization sectors. As a result, new
production systems are emerging, characterized by the prominent integration of data
and virtual technologies. These systems are so complex that they need model-based
management approaches.

M. K. Traore · S. Gorecki (B) · Y. Ducq

University of Bordeaux, CNRS, Bordeaux INP, IMS, UMR 5218, 33400 Talence, France
e-mail: [email protected]
M. K. Traore
e-mail: [email protected]
Y. Ducq
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 119
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_6
120 M. K. Traore et al.

The Digital Twin (DT) concept has risen as a pivotal strategy, landing in top
strategic technology trends. It revolves around the concept of mobilizing a model
in place of a system, ensuring continuous synchronization between the model and
the actual system. This synchronization serves to reflect real-world events on the
model, enabling the assessment of management initiatives on this ever-updated arti-
fact before implementation. Hence, the model is more than a simple representation,
evolving into a digital counterpart intricately linked to the specific system in question.
NASA pioneered the system-pairing approach, simulating situations in its space-
craft from distant command centers to guide astronauts in interventions. Notably, the
famous Apollo 13 mission in 1970 relied on a system-pairing technique but didn’t
involve a DT; instead, it utilized two Physical Twins (one in space and the other on
land).
The term “Digital Twin” was introduced by [1], acknowledging its underlying
principle foreseen by [2]. In the context of product lifecycle management, the con-
cept evolved through various stages: Mirrored Spaces Model [3], Information Mir-
roring Model [4], and finally as Digital Twin [5]. The concept is defined as “a set
of virtual information constructs that fully describe a potential or actual physical
manufactured product, from the micro-atomic to the macro-geometric level” [6].
This data-centric view shifted to behavioral aspects in [7], defining a DT as “an
integrated multi-physics, multiscale, probabilistic simulation of an as-built vehicle
or system that uses the best available physical models, sensor updates, fleet history,
etc., to mirror the life of its corresponding flying twin”. From the simulation per-
spective, this approach is disruptive, as simulation experiments are based on current
information from the system rather than assumptions. Simulation experiments rely
on the current information available from the system rather than assumptions [8, 9].
In this application, DT not only serves a representational function but also proves
effective in predicting the expected behavior of the system [10, 11]. Consequently, a
DT is not merely a comprehensive model of the system it represents but a collection
of integrated sub-models reflecting different aspects of the system [12]. Additional
dimensions have emerged, including the use of DT for prognostic and diagnostic
activities [13, 14], as well as real-time optimization [15, 16].
Nowadays, DT applications are diverse and extend across various industries,
including automotive [17], avionics [18], aerospace [7], energy [19], manufacturing
[8], health care [20], and services [21]. Within industrial contexts, DTs play a crucial
role in predictive maintenance of equipment, enhancing asset reliability and safety,
and optimizing both product design and process operations. In health care, the DT
approach holds the potential to revolutionize medical treatments by tailoring them
to the individual patient’s needs rather than relying on generalized approaches. Fur-
thermore, DTs contribute to the servitization trend by enabling companies to monitor
their products during customer usage, enhancing the overall customer experience.
DTs can be divided into three [22] main types: Digital Twin Prototype (DTP)
oriented for Product Lifecycle Management, Digital Twin Instance (DTI) oriented
for individual product that replicates one single physical or numerical ToI, and Digital
Twin Aggregate (DTA) that is an aggregate of several DTIs. In this chapter, we will
A Formal Framework for Digital Twin Modeling, Verification, and Validation 121

Fig. 1 High-level view of Digital Twin in interaction with both human and physical sides

mainly focus on DTI type, but contributions can be applied to both DTP, DTI, and
DTA.
Despite numerous research and development initiatives, the verification and vali-
dation (V&V) of DT remains a significant scientific challenge. To address this obsta-
cle, there is a pressing need for a comprehensive framework. This chapter responds
to this need by introducing a potential framework designed to achieve effective
DT verification and validation. As a second contribution, we propose three dif-
ferent approaches to reach the goal of DT V&V which can be applied to all DT
types: simulation based, software engineering based, and formal method based. This
approach extends to modeling, verifying, and validating the interoperability middle-
ware, specifically the Internet of Things (IoT) connection between the system and
its corresponding DT.
Thus, we can observe two main interacting elements: a physical or numerical
model of a system of interest, sometimes called Physical Twins, or Replica Twins
[6] in literature; however, we prefer to refer to it as the Twin of Interest (ToI),
enabling us to incorporate both physical models and non-physical ones, such as
numerical models. Perpetual synchronization between the TOI and the Digital Twin
is maintained, based on data tracked in the TOI. Considering that humans are in most
cases present in systems as stakeholders or decision-makers, e.g., health care [20],
smart cities [23], services [21], cyber physical and human systems [24], we can add
a third interacting element, the human.
Figure 1 provides an overview of the value chain, emphasizing key components
necessary for deploying an effective Digital Twin (highlighted in blue). The triangle
comprises three primary stakeholders: the physical (component to be replicated), the
digital (core Digital Twin), and the human (users and managers). On the physical
side, representing a cyber-physical system, the integration of cyber and physical
components allows operations executed by actuators, with data collected by sensors
and transmitted through a network. The digital side hosts the Digital Twin, receiving
data into models, enabling decisions sent back to the system or used by the user or
manager for further management decisions. The human side accommodates major
122 M. K. Traore et al.

stakeholders, including users and managers, utilizing the Digital Twin to explore use
cases and develop strategies.
Digital–physical interactions focus on the symbiotic relationship between the
physical part and the Digital Twin infrastructure, addressing scientific and techno-
logical challenges. These include designing the end-to-end data circuit, conceptualiz-
ing data needs, integrating heterogeneous data sources, and addressing requirements
for IoT liability and efficiency. Additionally, edge computing is employed for faster
computations related to real-time data streams, such as algorithms transforming raw
data for the Digital Twin’s use.
Digital–human interactions involve applying advanced technologies like Vir-
tual/Augmented Reality and Web/Mobile approaches to bridge the reality gap in
interfacing humans with the Digital Twin. This addresses issues like using Metaverse-
type technologies for the last mile to end-users, enhancing decision-making, and for-
malizing “cognitive interoperability” in the Metaverse through social interactions.
Digital-based human–physical interactions explore how Digital Twin models and
services impact the relationship between the Twin of Interest/physical subject and
the human subject: user (behavioral changes) or decision-maker.
The blue component of Fig. 1 (the Digital Twin part) can be zoomed in and illus-
trated as Fig. 2 according to an understanding framework highlighting the tree main
nodes-graphs common to all Digital Twins. The intertwining of these three nodes is
commonly referred to as a DT. However, a DT can be assimilated at different scales:
a process perspective (macro) or a product perspective (micro). For instance, in the
case of a production system, the DT at a process level could represent an entire indus-
trial production line. The feedback from the twin would include information about
the flows moving through the factory, with the twin providing guidance for managing
this chain. In the same context, at a different scale, from a “product” perspective,
the twin could be the DT of an engine in the production line. This distinction can
be illustrated with an example from a service system. The twin of a process (macro)
might be the DT of a smart city, while the twin of a product (micro) could be that of a
patient in the healthcare domain. This flexibility in scale allows the concept of Digital
Twins to be applied broadly, capturing both the holistic view of entire systems and
the detailed perspective of individual components or products within those systems.

2 Digital Twin Understanding Framework

In this chapter, we present a diagram illustrating the core DT. Figure 2 illustrates its
composition as three nodes: data, models, and services.
Figure 2 is the understanding framework, the symbiotic association of data, mod-
els, and services. In the Digital Twin, data stored and used are not only real-time
information from sensors but also draw from various legacy sources, aggregating
historical data, user equipment (e.g., smartphones, embedded cameras), open data,
and Internet-based APIs (e.g., Google Maps). Digital Twin models are structured into
modules, each focused to specific objectives and designed to address particular ques-
A Formal Framework for Digital Twin Modeling, Verification, and Validation 123

Fig. 2 High-level architecture for Digital Twin

tions (e.g., a physical system evolution in its environment under specific conditions,
the impact of adding/modifying infrastructures, identification of upcoming security
vulnerabilities, etc.). Modules are developed across specific domains depending on
the subject of interest.
As we can see on Fig. 3, one of the modules is the services provided by the Digital
Twin, also defined considering the needs and the area of the subject twinned. They
can be separated in four main categories: (1) monitoring data according to geometric
(2D/3D) or non-geometric models, or directly from data [25], (2) diagnosis is based
on regression, statistical, and stochastic models to identify and prevent events [26],
(3) prognosis capability [27] allows forecast or predicts future states, conditions, or
events based on the analysis of current and historical data and simulation models [7,
13], and (4) prediction ability to anticipate or forecast future outcomes, behaviors,
or trends based on simulation and optimization models [16].
Service node component is based on the model node which contains a collection
of integrated sub-models reflecting different aspects of the system twinned. Those
models can be geometric or non-geometric models for monitoring, statistical or
stochastic models for diagnosis services, simulation models and optimization models
for prognosis and prediction services. The use of these models fed by data enables
service execution. Moreover, these models can have the ability to update through
data–model interaction. Data node component serves as the foundational pillar that
includes all relevant information generated by real-time sensors, legacy data sources,
and data results from other services. The data node is crucial for fueling models;
thus, the quality and diversity of these data determine the reliability of predictions,
diagnostics, and analyses performed by the Digital Twin. By integrating varied data,
ranging from real-time conditions to historical data, the “Data” node provides a robust
foundation for prediction, diagnostics, and optimization services, contributing to an
informed and effective decision-making process.
124 M. K. Traore et al.

Fig. 3 DT models and services

Data–model integration stands out as a central challenge in the domain of Digital

Twin technology, representing a critical effort in ensuring the seamless adaptation of
data to dynamically detect changes in the real system. This integration aims at creating
a responsive environment where real-time data are assimilated on-the-fly to reflect
alterations of the physical system, enabling continuous updates to the corresponding
models [28]. The challenge takes place in establishing a harmonious relationship
between the evolving data stream and the digital models, ensuring that the virtual
representation remains accurate and aligned with the real-world counterpart.
Data–service integration is focused on services that can be directly derived from
raw data, bypassing the need for an intricate model. This involves organized access
to historical data for mainly monitoring purposes, allowing stakeholders to visu-
alize valuable information without relying on complex modeling structures. This
approach is an efficient way to access information for immediate decision-making
and monitoring activities.
Model–service integration takes a different approach by focusing on how models
are employed and potentially combined to feed services. This integration is pivotal
in optimizing the utility of digital models, exploring their versatility in providing the
largest spectrum of services possible. By strategically aligning models with specific
services, this integration ensures that the Digital Twin leverages its modeling capabil-
A Formal Framework for Digital Twin Modeling, Verification, and Validation 125

Fig. 4 Digital, physical, and human interoperations

ities to offer a diverse range of functionalities, contributing to its overall effectiveness

in decision support and system optimization.
The three main nodes composing the DT depicted in Fig. 3 will be in continuous
interaction with external entities through various means, protocols, and at varying
frequencies. Figure 4 provides a detailed breakdown of the elements interacting with
the DT and the nature of the interactions it will exchange with them. The intercon-
nectedness is established through a network of dynamic exchanges, emphasizing the
diverse and multifaceted engagements the DT maintains with external entities.
Figure 4 serves as a comprehensive synthesis of insights from both Figs. 1 and 2.
It introduces a nuanced perspective on the human side, distinguishing between the
user and the decision-maker. To illustrate the intricacies of this figure, we’ll delve
into a case study focusing on a smart city. The case of smart cities illustrates this
high-level architecture; however, it’s just an example that can be swapped with other
Digital Twins where humans are in the loop (e.g., DT health care). In this smart city
scenario, the Twin of Interest (ToI) is the city itself, intricately twinned with a DT
(hairy circle at the bottom of Fig. 3). This digital representation engages in continuous
interaction with the “Human” system, segmented into two pivotal sub-systems: the
decision-maker, embodying the smart city manager, and the user, encompassing all
other inhabitants and stakeholders within the smart city, such as citizens.
126 M. K. Traore et al.

Moreover, the complexity shown in the figure includes various layers of inter-
actions involved in the Digital Twin’s overall process. The data node of the Digital
Twin emerges as the gateway facilitating seamless communication with the physical
realm. Concurrently, the services’ node acts as the portal enabling the commissioning
(activation and utilization) of the Digital Twin by the urban decision-maker. On the
other hand, citizen engagement finds its avenue through the models’ node, shaping
a dynamic and interactive ecosystem within the smart city’s digital representation.
This intricate network of nodes and interactions highlights the sophisticated web of
relationships and functionalities within the Digital Twin framework.
Having explored the structure and high-level architecture of the Digital Twin in
this chapter, the next logical step is to dive into the modeling and formal specifi-
cation aspects. Understanding the composition and intricacies of the Digital Twin
sets the stage for a more detailed examination of its modeling and standardization in
the following chapter. This transition is pivotal as it emphasizes the need for a sys-
tematic and formalized approach to enable the subsequent phases of verification and
validation. With a comprehensive understanding of the Digital Twin’s composition,
the focus now shifts toward establishing robust models and standardized practices,
paving the way for a thorough examination of its capabilities and functionalities.

3 Digital Twin Modeling

Formal specification brings valuable advantages. Firstly, it clarifies the intended con-
cept, making it less open to different interpretations. Secondly, it allows the creation
of systematic approaches to generate models that can be executed. Thirdly, it pro-
vides a possibility for symbolic manipulation to formally check the consistency of the
specification and its alignment with certain requirements: verification and validation
(V&V). Thus, the need for methods and tools for modeling and specifying Digital
Twins is a crucial point addressed in this document. Without a formal description of
these twins, it becomes impossible to perform formal verification and specification
to ensure their alignment with the case study, regardless of the application domain.
The scientific community agrees that there is a significant difference between tra-
ditional models and digital twins. In [29–32], authors provided summaries of differ-
ent definitions of Digital Twins and proposed generalized definitions and distinctive
characteristics to differentiate Digital Twins from other models.
However, it is challenging to find traces of research that focus not on modeling
Digital Twin models, but on modeling the Digital Twin itself. Some attempts have
been made to achieve this goal, but they are often not sufficiently thorough. In [33],
authors identified and categorized different models within Digital Twins into six
categories: application domain, hierarchy, discipline, dimension, universality, and
functionality. However, this contribution mainly focuses on analyzing models within
Digital Twins, without addressing the overall model of Digital Twins itself. The same
author, in [34], proposes to deconstruct and investigate the digital models into six
distinct aspects: model construction, model assembly, model fusion, model verifi-
A Formal Framework for Digital Twin Modeling, Verification, and Validation 127

cation, model modification, model management, but once again, this work focuses
on the models inside the Digital Twin, not the general model of the DT itself. A
similar work has been done in [35] where author in this paper reviews and analyzes
several methods and modeling techniques used in DT domain in order to classify
modeling methods in groups, e.g., physics-informed ML; data-driven modeling; sys-
tem modeling; physics-based modeling; geometric modeling, etc. We will see that
this approach brings an interesting perspective because depending on the application
domain, modeling methods differ, which implies variations in verification and vali-
dation methods. In [22], a DT is presented as a composition of basic components that
provide on one hand, basic functionalities (identification, storage, communication,
security, etc.), and on the other hand, an aggregated DT is defined as a hierarchical
composition of other DTs, which is a first trace in the literature of a model not internal
to the DT, but of the DT itself. The article proposes a reference architecture of DT
using Automation ML through model-driven engineering approach.
Based on the state of the art, we can observe that DT modeling domain covers
two main aspects. First aspect is “modeling for DT", which means the design of the
model within the Digital Twin, or the architecture of the Digital Twin. The second
aspect is “modeling of the DT itself”, where there are very few articles, if any, that
talk about modeling the Digital Twin itself.
Formal specification methods offer significant advantages for Digital Twin mod-
eling [36], regardless of the application context. However, as we can see, there is
no standardized approach for the formal modeling of Digital Twins. Consequently,
facilitating the verification and validation (V&V) of Digital Twins is not a straight-
forward task. This gap in standardization shows the need for a systematic and formal
framework that can be universally applied to ensure the accuracy and reliability of
Digital Twins across diverse domains. Establishing such a standard would contribute
to advancing the field, providing a common ground for researchers and practitioners
to enhance the V&V processes and overall robustness of Digital Twin implementa-
tions.
To formally specify a DT, we use a system-theoretic approach introduced in 1976.
In this approach [37], the DT model is considered a black box with inputs and outputs
that connect it to the environment through sensors and actuators. Zeigler’s paradigm
includes fundamental entities in the modeling and simulation field. These entities
are represented by Fig. 5 left side, consisting of the source system (the system under
study), the context, the model, and the experimental frame (EF). Please note that the
simulator is not shown here. The source system provides behavioral data, the context
defines the observation conditions, the model represents the system abstractly using
rules or equations, and the EF, acting as an abstract representation of the context,
couples with the system model to generate data under specific conditions. The sim-
ulator, although not illustrated, is the automaton capable of executing instructions
from the resulting coupled model.
In contrast to a conventional model (left side in Fig. 5), which is built as an
abstract representation of a source system for experimentation under specific condi-
tions defined by an Experimental Frame (EF), a DT model (right side in Fig. 5) is
continuously refreshed with data from the source system. This ongoing update neces-
128 M. K. Traore et al.

Fig. 5 DT system-theoretic approach

sitates an inference mechanism (see data–model integration of Fig. 2) to maintain the

model’s validity within the same EF. This formalization is crucial for ensuring the
accuracy and reliability of the DTs through the inference-model mechanism, but
raises issues related to verification & validation techniques.
Introducing the Discrete Event System Specification (DEVS) paradigm, [37]
defines a model as a mathematical and logical object which serves as a way of
specifying a dynamic system, with the following elements:
• State: A state can either be a “hold state” or a “passive state”. A hold state is
one that the model will stay in for a certain amount of time before automatically
changing to another state (via an internal transition). A passive state is one that the
model will remain in indefinitely (or until it receives a message that triggers an
external transition). One state in the model must be designated as the initial state
from which all interactions with the external word commence.
• Time Advance: Every state has a time advance value which specifies the amount of
time that expires before it automatically changes to another state (via an internal
transition). The time advance for a “hold” state is a finite real value. The time
advance for a “passive” state is infinity.
• Internal Transitions and Output: Every hold state in the model has one internal
transition defined in order to specify the state to which the model should transition
after the specified amount of time. Any state that has an internal transition can
also have one or several output messages that are generated before that internal
transition occurs.
• Input and External Transitions: Any state can have one or more external transitions
defined. An external transition defines an input message that the model might
receive when in a given state and the state to which the model should transition in
reaction to that input message.
Our framework extends these elements with the following:
• Phase: In the case of a discrete event system, a phase is a state, i.e., its properties
are defined by a vector of discrete values. In the case of a hybrid (i.e., combined
discrete-continuous) system, the phase properties are defined by equations (linear,
A Formal Framework for Digital Twin Modeling, Verification, and Validation 129

differential, etc.). Examples of phase for a vehicle are “accelerating”, “cruise-

controlling”, “decelerating”, “stopping”, etc., where each phase can be defined by
some equations binding the position, speed, acceleration rate, etc., of the vehicle.
A phase can be seen as a set of states logically grouped together (and formalized
by the properties defining that phase).
• Semantic domain: A phase is defined by its properties, which are expressed by
means of state variables. These variables form the semantic domain of the vehicle
phases. For example, position, speed acceleration rate, etc., form the semantic
domain of the vehicle previously mentioned.
• Parameter: A parameter is an abstraction of some assumption or constraint on
the reality represented by the model. Usually, the value of a parameter doesn’t
change during a simulation experiment, as the experiment is performed under
some circumstances that the value of the parameter captures (partially or entirely).
The definition of this value for each of the parameters of a model is what is
called model calibration (or tuning). However, simulation experiments with a DT
model must be possible under varying circumstances, which requires the dynamic
change of value of a parameter (i.e., dynamic calibration) be possible during the
same experiment.
• Activity: An activity is a set of computations performed by a model during a phase
or during the transition from one phase to another. The duration of any activity
is zero in the simulated time (it is an instantaneous operation in the real world).
Activities are formalized with logical predicates.
We then specify a DT as a 7-uplets . M =< Δ, X, Y, Φ, ∑, Δ, Δ∗ >, where:

• .Δ is the parameter set: It models all assumptions made on the context in which the
DT model is used. .∀ λ ∈ Δ, dom(λ) is the set of all admissible values for .λ. .Δ is
defined by variables which values can be updated by the context sensors pairing
the DT with the ToI.
• . X is the input set: Its models the influences received from the DT environment
.∀ x ∈ X, dom(x) is the set of all admissible values for x.

• .Y is the output set; it models the DT influences on the environment. .∀y ∈

Y, dom(y) is the set of all admissible values for y. X and Y are defined by variables
whose values can be updated by the interface sensors pairing the DT with the ToI.
• .Φ is the phase set; it models the stable discrete/continuous steps of the DT model.
Each phase is defined by its properties, as well as the amount of time that expires
before it changes to another phase (via an internal transition), and the activities
performed during the phase.
• .∀ ϕ ∈ Φ, Φφ =< πϕ , θϕ , τϕ > is the phase’s definition, with:

– .πϕ is the phase properties, a predicate on .∑ (noted .πϕ ∈ ℘ (∑) ∪ N I L) giving

the semantics of .π in .∑.
– .θϕ is the phase activitu, a predicate on .∑ modeling the computations performed
during .ϕ.
– .τϕ is the phase’s time advance, a delay (i.e., expected lifetime of .ϕ; τϕ ∈ R+ )
before the model changes from to another phase.
130 M. K. Traore et al.

• .∑ is the semantic domain: This is the set of variables, which the phases are mapped
onto. .∀ σ ∈ ∑, dom(σ ) is the set of all admissible values for .σ . .∑ is defined by
variables which values can be updated by the system sensors pairing the DT with
the ToI.
j j! j! j j!
• .Δ = {Δi? , Δi , Δi? , i ∈ Φ, j ∈ Φ} is the phase-to-phase transition set. .Δi? , Δi ,
j!
and .Δi? are respectively external, internal, and confluent transitions, from phase
to phase (an internal transition occurs when the delay of the current phase elapses;
an external transition is triggered by the receipt of an input; a confluent transition
occurs when both internal and external transition conditions occur simultaneously),
with:
j
– .Δi? =< ωi?, j , εi?, j , θi?, j > where:

· .ωi?, j ∈ ℘ (X ) is the condition of input receipt.

· .εi?, j ∈ ℘ (∑) ∪ {N I L} is the condition of transition.
· .θi?, j ∈ ℘ (∑) ∪ {N I L} is the action performed in transition.

j!
– .Δi =< εi, j! , ρi, j! , θi?, j > where :

· .εi, j! ∈ ℘ (∑) ∪ {N I L} is the condition of transition.

· .ρi, j! ∈ ℘ (∑) ∪ {N I L} is the output sending predicate.
· .θi, j! ∈ ℘ (∑) ∪ {N I L} is the action performed in transition.

j!
– .Δi? =< ωi?, j! , εi?, j! , ρi?, j! , θi?, j! > where:

· ε
. i?, j! ∈ ℘ (X ) is the condition of input receipt.
· ε
. i?, j! ∈ ℘ (∑) ∪ {N I L} is the condition of transition.
· .ρi?, j! ∈ ℘ (∑) ∪ {N I L} is the output predicate.
· .θi?, j! ∈ ℘ (∑) ∪ {N I L} is the action done in transition.

– .Δ∗ =< i ∗ , θ ∗ , τ ∗ > is the initialization, giving the phase and action prior to the
DT execution, where :

· .i ∗ the initial phase.

· .θ ∗ is the activity performed in initializing the model.
· .τ ∗ is the time already elapsed in the initial phase.

As an example, let’s consider the DT of a smart city which operates a ring road
management system (Fig. 6). The DT collects data on weather and congestion through
road sensors and controls speed limit of the highway. Figure 4 illustrates the compo-
nents involved in the engineering of this DT (Fig. 7).
The DT model of a ring road management system presents three phases: stopped
(i.e., accident or roadwork), regular (i.e., good weather conditions), and free-flowing
A Formal Framework for Digital Twin Modeling, Verification, and Validation 131

Fig. 6 Digital Twin of a smart city—road management system

Fig. 7 DT formal specification metamodel

traffic. The quantity of vehicle (vehicle flow) will be noted F and the speed limit
(determined by the system) will be noted “S”. Roads’ number will be “R” and
number of vehicles “n”. Weather factor W), and reduced speed limit (i.e., due to bad
weather conditions or traffic jams).
The formal specification of this model is illustrated by 8 and given by:
DT Road Management = .< Δ, X, Y, Φ, ∑, Δ, Δ∗ >, where:
• .Δ = {n, W }, .n and .W can be updated with context data (number of vehicles and
weather factor).
• . X = {in, war ning} and .Y = {F}, in can be updated with interface data (i.e.,
modifying speed limit), and the speed limit set is sent through the output pin.
132 M. K. Traore et al.

• .∑ = {R, F}, R (number of roads) can be updated with system data (i.e., object
fallen on a road), and . F (speed limit) is calculated, but the model is in regular or
reduced phases.
• .Φ = {stopped, r egular, r educed}, with

– .Φstopped =< F = 0, N I L , +∞ >, the road will remain stopped until

switched on (speed limit positive). No activity on this phase, as indicated by
NIL.
– .Φr egular =< F = 130, N I L , +∞ >, the road is open, and speed limit is
cruising for a infinity time.
– .Φr educed =< F = f (n, W ) × 110, N I L , +∞ >, the road is open but speed
limit is reduced by 25% for a infinity time.
r egular stopped r egular
• .Δ = {Δstopped? , Δr egular ? , Δrr egular
educed
? , Δr educed? with:

r egular
– .Δstopped? =< in?on, N I L , N I L >, the external transition from stopped phase
to regular phase arises when road manager switches it on, without any other
condition (as indicated by the first NIL) or activity (as indicated by the second
NIL).
stopped
– .Δr egular? =< in?o f f, N I L , N I L >, the external transition from regular phase
to stopped phase arises when the road manager switches it off.
– .Δrr egular?
educed
=< war ning?o f f, N I L , I = W >, the internal transition from reg-
ular phase to reduced phase arises when the weather factor (W) crosses a thresh-
old.
r egular
– .Δr educed? =< war ning?on, N I L , I = W >, the internal transition from
reduced to regular phase arises when the weather factor (W) crosses again the
threshold.

• .Δ∗ =< r egular, N I L , 0 >, the speed limit is initially regular (before any simu-
lation experiment).

The adoption of this formal approach allows the DT developers to clarify the
intended concept behind each model that composes DT and allows possibility for
symbolic manipulation in order to formally check the consistency through several
verification and validation methods. The DEVs’ model we have just described pro-
vides a formal framework for structuring the “model” component (as depicted in
Fig. 2) of the Digital Twin. This formalism can be extended to rigorously describe
the various services that may interact with both the model and data layers. Adopt-
ing a consistent approach for specifying and understanding the intricate dynamics of
services within the Digital Twin ecosystem is essential for ensuring seamless interop-
erability and functionality. The utilization of such formalism will play a pivotal role
in achieving a comprehensive and well-defined representation of the entire Digital
Twin framework. This enables a more detailed analysis and validation of the intercon-
nected components, fostering a deeper understanding of the intricate relationships
between services, models, and data layers within the Digital Twin paradigm.
A Formal Framework for Digital Twin Modeling, Verification, and Validation 133

However, the “DT Model Inference” layer is not explicitly addressed in the for-
malism proposed by Zeigler [37] (Fig. 5 left side). It remains an open challenge for
the scientific community to establish a method for formalizing the data inference over
model within the Digital Twin framework. A DEVS model can adeptly represent and
design various types of models, such as multi-agents, discrete events, process flows,
and cellular automata. Moreover, a DEVS model can also represent several services,
such as prediction, diagnosis, and optimization services. The scientific community
must now engage in defining a formalism that seamlessly accommodates the repre-
sentation of a data lake, regardless of its form (whether it be a database, compilation
of data sources, files, a centralized data entry point, ontology, etc.). This objective
necessitates a versatile approach that aligns with the diverse nature of data repre-
sentation in different Digital Twin applications, thereby ensuring a comprehensive
and adaptable framework for handling the intricacies of data management within the
Digital Twin paradigm.

4 Digital Twin Verification and Validation Methodology

Verification and validation is indispensable in ensuring the accuracy and reliability

of Digital Twins. It must be based on a clear and precise specification formalism [38],
especially in the context of increasingly complex systems such as smart cities. In the
literature, many works mention the verification and validation of various systems, but
there are very few documents that delve into this issue regarding Digital Twins [39].
Among these papers, various approaches have been used in recent research. These
approaches can be broadly classified into two main categories: offline and online.
Offline validation of a Digital Twin model involves providing it with historical data
to evaluate its accuracy and calibrate its parameters accordingly. In contrast, online
validation continuously feeds real-time data from the physical system to the Digital
Twin model, allowing for dynamic calibration of the Digital Twin model’s parameters
to maintain its accuracy and adaptability. Research on the verification and validation
of Digital Twins is still in its early stages, and only a limited number of comprehensive
studies have been conducted. Most existing studies have focused on offline validation
methods to evaluate Digital Twin models: Khan et al. [40] proposed an offline model-
based testing approach to validate the model. The proposed approach uses a modeling
tool to create a specification model and generate test cases for a production system.
Hasan et al. [41] proposed a predictive Digital Twin model for autonomous surface
vessels to diagnose faults. Argota Sánchez-Vaquerizo [42] addressed the issue of
balancing heavy traffic flow in a complex urban network using a Digital Twin. Finally,
other works exist [43] providing a state-of-the-art review of V&V of cyber-physical
systems by classifying existing V&V methods in this field, while emphasizing that
they are currently insufficient.
Based on what we just observe, there is no notable work on the verification &
validation of Digital Twins. However, we see three ways of approaching the DT V
& V issue:
134 M. K. Traore et al.

• V&V the Digital Twin infrastructure by software engineering [44]. Considered

as a system. Verification and validation methods from systems engineering are
applied. V&V involves unit validation, followed by integration validation, and
then system validation: Sect. 4.1.
• V&V by Modeling and Simulation [45]. This involves constructing a simulation
model of the DT using the DEVS approach: Sect. 4.2.
• DT analysis by formal method [46] allows V&V if the DT has certain properties
and checking if it does not have certain properties, e.g., constructing the possibility
tree of the DT, or use formal simulation techniques: Sect. 4.3.

4.1 V&V by Software Engineering

The mechanisms of V&V for complex systems like Digital Twins can draw inspi-
ration from methods within the field of software engineering. It is well known that
software engineering uses several levels of tests to verify and validate any software
system [47]. These methods can be applied on each types of DT (DTP, DTI, and
DTA) dependent on the test level chosen. From a software engineering point of view,
V&V can be broken down into clearly identified stages that we propose to map
with DT architecture and described in Fig. 8: iterative unit tests (I), initially focus-
ing at the lowest level of testing. Here, each sub-components (including the three
main entities I.1-I.2-I.4 and inference model interaction I.3) are tested locally. The
goal of this testing level is to insure that the component being tested is conforming
to its specifications and ready to be integrated with other components of the DT.
The second level is integration testing (II of Fig. 8) which consists in insuring that
the interfaces between components are correct and the sub-components combined
to execute the DT’s functionality correctly. It goal is to assess the communication
between Twin of Interest (II.1) and the entire Digital Twin (II.2). Finally, the last
level is system testing, which is the process of verifying that the entire DT meets its
specified requirements (the all structure in Fig. 8).

4.1.1 Unit Test

Unit tests involve verifying and validating the smallest building blocks of the system.
The aim is to meet the requirements of each sub-component of the twin. In the frame-
work case, this involves addressing the data, models, and services’ sub-components
of Fig. 8. These unit tests are at the microscopic scale, thus corresponding to the tests
of DTI (Digital Twin Instance) type.

• Methods for Models’ verification and validation (I.1 of Fig. 8) are essential to
ensure the accuracy and reliability of models within DTs. Different types of mod-
els exist for different fields of application [48] (physical, mathematical, statistical,
A Formal Framework for Digital Twin Modeling, Verification, and Validation 135

Fig. 8 V&V unit tests (I) and integration tests (II)

conceptual, 3D graphical, IT, economic, simulation, data, process, systemic mod-

els, etc.). There are as many V&V methods as there are formalisms [49]. Many
works exist in the literature about model V&V methods, of all possible types
[49]. Most known methods used for it are: residual method [50], cross-validation
[51], statistical analysis (regression analysis [52], variance analysis [53], hypothe-
sis testing [54]), stability test [55], cross-model validation [56], metric evaluation
[57], mean square error [58], etc.). By combining these methods and techniques
according to the context of the targeted Digital Twin, it is possible to guarantee the
adequacy of the models with the Physical Twin, and therefore, with the physical
object.
• Methods for Data verification and validation (I.2 of Fig. 8) are essential to ensure
the ability to carry a given type of traffic under the right conditions, in terms of
availability, throughput, transmission times, quality, completeness, value, avail-
ability, etc. [59]. The importance of data quality has been highlighted in various
aspects of operating processes [60], decision-making activities [61], and inter-
organizational cooperation requirements [62]. Both public and private sectors have
initiated numerous efforts where data quality plays a central role. For instance, the
UC government passed the Data Quality Act in 2002 [63], while the Government
of USA introduced the Data Quality Initiative Framework in 2004 [64]. Therefore,
formal methods exist in the literature [65] to verify that a data model adequately
meets the requirements set for it based on its application domain: Quality Of Ser-
vices (QOS). According to the literature, most of these methods can be classified
into different techniques: data-driven techniques (e.g., [66] or [67]); process-driven
techniques (e.g., [68] or [69]).
136 M. K. Traore et al.

Those techniques of verifying the QOS of a dataset can be applied on a DT as unit

testing for the data lake layer.
– Methods for Services’ verification and validation (I.3 of Fig. 8). Before discussing
V&V of services, it is important to establish boundaries on what we will call
“Services”. In any domains (even outside of DTs), each service relies on both data
and models. This is also the case in our context of Digital Twins. There is a wide
variety of services that can vary greatly depending on the application domains. In
our case study, and to encompass the broadest spectrum of Digital Twins possible,
we define several categories of services:

• Monitoring services, for data visualization: relies on statistical models or visual-

ization models. Two types of monitoring services can be distinguished; e.g., 2/3D
geometric representation or non-geometric (e.g., statistical models);
• Diagnostic services (limited to “diagnosis models”): relies on regression, statisti-
cal, or stochastic models;
• Prognostic/predictive services, which rely on simulation models (discrete event,
continuous, or hybrid), regression models, optimization (discrete or continuous);
• Prescription services: based on simulation or optimization models to enable
decision-making, control, or feedback from a Digital Twin.
• Control services, which allow to continuously maintain system in a desired state
by providing adapted responses.
Based on literature, V&V of DTs services is not a significant problem as numerous
works exist to address this issue. Services within the Digital Twin can be V&V
by classical QOS methods [70]. We also can take inspiration from V&V of Services
Oriented Architecture [71], or from web-services framework V&V applications [72].

4.1.2 Integration Tests

In software engineering, integration testing is a testing phase aimed at verifying that

the various modules or components of a system interact correctly when combined
to form subsystems or the complete system. Integration testing is at the macro-level
of software engineering test, and it can be associated to Digital Twin aggregate level
V&V. The main objective of integration testing is to detect and resolve any potential
interaction problems between modules, such as interferences, incompatibilities, or
communication errors. Integration tests can be conducted at different levels, including
the level of individual components, subsystems, or the entire system.
• Interaction between layers within the Digital Twin (I.3 of Fig. 8) is part of the
interaction testing. The interactions between the layers of the DT (model, services,
data) must also be part of the unit testing and V&V process. For this, the difficulty
will be unequal. Communication between services and models is a trivial task as
this configuration is common in many cases. Similarly, communication between
data and services is straightforward. However, a much more complicated part
A Formal Framework for Digital Twin Modeling, Verification, and Validation 137

lies in the interactions between data and models. Here, the context of the Digital
Twin imposes a very different treatment compared to other contexts. Thus, a data
inference mechanism takes place in the DT. Data can, on the fly, influence the
structures of the models [73]. It is therefore important to note that this part of the
V&V opens up prospects for evolution in the scientific community.
• Integration testing between the ToI and data layer (II.1 of Fig. 8) is crucial to ensure
seamless interaction and functionality. This integration testing phase focuses on
verifying that the ToI effectively communicates with the data layer and that the
data exchanged between them are accurate, consistent, and appropriately utilized.
Testing scenarios may include validating data input/output mechanisms, ensuring
compatibility between data formats and structures, and verifying the synchroniza-
tion of data updates between the ToI and the Digital Twin’s data layer. Additionally,
integration testing evaluates the responsiveness of the Digital Twin to queries and
commands from the ToI, ensuring that the desired actions are executed correctly
and in a timely manner. Through rigorous integration testing, developers can iden-
tify and address any potential issues or discrepancies in the interaction between
the Twin of Interest and the data layer, ultimately enhancing the reliability and
effectiveness of the Digital Twin system.
• Entire Digital Twin (II.2 of Fig. 8) testing phase focuses on validating the inte-
gration points between the three layers to ensure that they work harmoniously to
achieve the intended objectives of the Digital Twin. Testing scenarios may include
verifying the flow of data between the data layer and the model layer, ensuring
that data inputs are accurately processed by the models and that the outputs are
consistent with expected results. Additionally, integration testing evaluates the
interaction between the model layer and the services layer, ensuring that services
can effectively leverage the models to provide desired functionalities. Moreover,
integration testing examines the integration points between the data layer and the
services layer, validating that services can access and utilize data from the data layer
efficiently. By thoroughly testing these integration points, developers can identify
and address any potential issues or inconsistencies in the interactions between the
layers of the Digital Twin, ultimately ensuring the reliability and effectiveness of
the entire system.

4.1.3 System Test

System testing serves as a crucial phase aimed at evaluating the overall functionality,
performance, and behavior of the Digital Twin as a cohesive system. System testing
for a Digital Twin involves assessing its ability to fulfill its intended purpose and
objectives across various scenarios and use cases. This scale level can be associated
to Digital Twin Prototype. The comprehensive testing process examines the interac-
tion and integration of all components within the Digital Twin, including the data
layer, model layer, and services layer. It verifies the accuracy of data inputs, the effec-
tiveness of models in simulating real-world behavior, and the reliability of services
138 M. K. Traore et al.

in leveraging the Digital Twin for decision-making and optimization. System testing
also encompasses testing under different environmental conditions and stress levels
to ensure the robustness and resilience of the Digital Twin. By conducting system-
atic and rigorous system testing, developers can identify any potential weaknesses,
errors, or performance bottlenecks in the Digital Twin and take corrective actions to
enhance its overall quality and reliability.

4.2 V&V by Modeling and Simulation

Simulation-based verification and validation is an approach used to assess and ensure

the quality, reliability, and accuracy of simulation models. This method relies on the
use of computer simulations to verify and validate models by comparing them to
real data or reference standards [74]. In the context of simulation-based V&V, the
real system is replaced by a simulation model. Thus, the Digital Twin is subjected
to various tests and scenarios to evaluate its performance and behavior under varied
conditions, replicating cases that may occur in reality. This may include sensitivity
tests to assess the model’s response to variations in input parameters, robustness
tests to evaluate its ability to produce consistent results under varied conditions,
and validation tests to compare simulation results to experimental data or theoretical
results. The primary objective of this method (Model Continuity) is to ensure that the
Digital Twin (both DTI and DTA) is capable of adapting and accurately reproducing
the behavior of the real system it represents and providing reliable and meaningful
results to support decision-making and understanding of the studied phenomenon
[45].

4.3 V&V by Formal Method

Formal methods represent a powerful alternative for carrying out the V&V of Digital
Twin Instance. They provide structured and systematic techniques to ensure the
compliance of Digital Twins with the specifications and requirements of the real-
world system.
Formal methods are approaches [75] based on mathematical and logical concepts
to specify, design, and verify software or hardware systems. They rely on formal
languages and formal analysis techniques to guarantee the correctness and safety
of systems. In the context of Digital Twins, formal methods can be used to specify
expected behaviors, detect design errors, and validate the consistency between the
virtual model and the real system.
The first step in V&V of Digital Twins using formal methods is to develop a
precise formal modeling of the system. This modeling can be carried out using
formal languages such as Petri nets [76], automata [77], or formal specification
A Formal Framework for Digital Twin Modeling, Verification, and Validation 139

languages. By formally describing the system’s behavior and its interactions with the
environment, this modeling provides a solid foundation for analysis and verification.
Once the formal model is established, various techniques of formal verification
can be used to ensure its correctness and compliance with specifications. Among
these techniques, model checking [78] involves automatically verifying if the model
satisfies certain specified properties by traversing the possibility tree to determine
the absence of deadlock or system termination. Theorem proving techniques [79], on
the other hand, use mathematical reasoning methods to demonstrate the correctness
of the model against a predefined set of properties.
In addition to formal verification, formal validation of Digital Twins aims to ensure
that the virtual model accurately reflects the behavior of the real system. This step
may involve the use of formal simulation techniques [80], where the formal model
is executed in a simulation environment to evaluate its behavior under different
conditions. Comparing the results of formal simulation with real data validates the
accuracy and reliability of the Digital Twin.
Formal methods offer a systematic and rigorous approach to the verification and
validation of systems, characteristics that can be beneficial for the Digital Twin
domain. By using formal languages, advanced analysis techniques, and specialized
tools, it is possible to ensure the reliability and accuracy of these virtual models,
paving the way for their use in critical applications such as industry, health care, and
transportation.

5 Conclusion

In this paper, our first contribution lies in proposing the design of an architecture
for understanding the Digital Twin, enabling high-level modeling. This architectural
understanding framework provides a structured approach to conceptualizing and
representing Digital Twins, facilitating their effective modeling and analysis through
three main levels: data level, model level, and service level.
Building upon this foundation, our subsequent analysis delves into the question
of verification and validation of Digital Twins, a topic that has received limited atten-
tion in the existing literature despite the increasing prominence of this technology.
Recognizing this research gap, we have conducted an in-depth analysis to explore
the potential of existing methods in the state of the art to establish the conceptual
and methodological foundations for ensuring the accuracy and reliability of Digital
Twins. Our investigation has identified three primary axes: software engineering,
modeling and simulation, and formal methods, each offering valuable insights into
addressing the verification and validation challenges associated with Digital Twins.
This study serves as both a comprehensive review of the current state of research
and a roadmap for future endeavors in this emerging field. By consolidating existing
knowledge and pointing areas requiring further investigation, our work aims to facil-
itate significant progress in the verification and validation of Digital Twins, fostering
their widespread adoption across diverse application domains. Through collabora-
140 M. K. Traore et al.

tive efforts and continued research, we envision a future where Digital Twins play
an increasingly integral role in enhancing decision-making processes and driving
innovation across various industries.

References

1. Piascik RS et al (2010) DRAFT materials, structures, mechanical systems and man-

ufacturing roadmap-technology Area 12”. In: Obtained from: http://www.nasa.gov/pdf/
501625mainTA12-MSMSM-DRAFT-Nov2010-A.pdf
2. Gelernter D (1993) Mirror worlds: or the day software puts the universe in a shoebox... How
it will happen and what it will mean. Oxford University Press
3. Grieves MW (2005) Product lifecycle management: the new paradigm for enterprises. Int J
Product Dev 2(1–2):71–84
4. Grieves M (2006) Product lifecycle management: driving the next generation of lean thinking
5. Grieves M (2011) Virtually perfect: driving innovative and lean products through product
lifecycle management, vol 11. Space Coast Press Cocoa Beach
6. Grieves M, Vickers J (2017) Digital twin: mitigating unpredictable, undesirable emergent
behavior in complex systems. In: Transdisciplinary perspectives on complex systems: new
findings and approaches, pp 85–113
7. Glaessgen E, Stargel D (2012) The digital twin paradigm for future NASA and US Air Force
vehicles. In: 53rd AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics and materials
conference 20th AIAA/ASME/AHS adaptive structures conference 14th AIAA, p 1818
8. Rosen R et al (2015) About the importance of autonomy and digital twins for the future of
manufacturing. Ifac-papersonline 48(3):567–572
9. Grieves MW (2019) Virtually intelligent product systems: digital and physical twins
10. Boschert S, Rosen R (2016) Digital twin—the simulation aspect. In: Mechatronic futures:
challenges and solutions for mechatronic systems and their designers, pp 59–74
11. Michael S, Juergen R (2016) From simulation to experimentable digital twins: simulation-based
development and operation of complex technical systems. In: IEEE international symposium
on systems engineering (ISSE). IEEE, pp 1–6
12. Elisa N, Luca F, Marco M (2017) A review of the roles of digital twin in CPS-based production
systems. Proc Manuf 11:939–948
13. Reifsnider K, Majumdar P (2013) Multiphysics stimulated simulation digital twin methods for
fleet management. In: 54th AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics,
and materials conference, p 1578
14. Fei T et al (2018) Digital twin-driven product design, manufacturing and service with big data.
Int J Adv Manuf Technol 94:3563–3576
15. Söderberg R et al (2017) Toward a Digital Twin for real-time geometry assurance in individu-
alized production. CIRP Ann 66(1):137–140
16. Hao Z et al (2017) A digital twin-based approach for designing and multi-objective optimization
of hollow glass production line. IEEE Access 5:26901–26911
17. Violeta Damjanovic-Behrendt (2018) A digital twin-based privacy enhancement mechanism
for the automotive industry. In: 2018 international conference on Intelligent Systems (IS).
IEEE, pp 272–279
18. Tuegel EJ et al (2011) Reengineering aircraft structural life prediction using a digital twin. Int
J Aerosp Eng
19. Zhang M, Zuo Y, Tao F (2018) Equipment energy consumption management in digital twin
shop-floor: a framework and potential applications. In: 2018 IEEE 15th International Confer-
ence on Networking, Sensing and Control (ICNSC). IEEE, pp 1–5
20. Bramlet M et al (2016) Virtual reality visualization of patient specific heart model. J Cardiovasc
Magnetic Resonance 18(1):1–2
A Formal Framework for Digital Twin Modeling, Verification, and Validation 141

21. Bolton RN et al (2018) Customer experience challenges: bringing together digital, physical
and social realms. J Service Manage 29(5):776–808
22. Schroeder GN et al (2021) A methodology for digital twin modeling and deployment for
industry 4.0. In: Proceedings of the IEEE, vol 109, issue 4, pp 556–567. https://doi.org/10.
1109/JPROC.2020.3032444.
23. White G et al (2021) A digital twin smart city for citizen feedback. Cities 110:103064
24. Barnabas J, Raj P (2020) The human body: a digital twin of the cyber physical systems. Adv
Comput 117(1):219–246
25. Bailenson JN, Segovia KY (2010) Virtual doppelgangers: psychological effects of avatars who
ignore their owners. In: Online worlds: convergence of the real and the virtual, pp 175–186
26. Jain P et al (2019) A digital twin approach for fault diagnosis in distributed photovoltaic
systems. IEEE Trans Power Electronics 35(1):940–956
27. Zhou X et al (2023) A fuzzy-set-based joint distribution adaptation method for regression and
its application to online damage quantification for structural digital twin. Mech Syst Signal
Process 191:110164
28. Mama D, Kaba TM (2023) Formal approach to digital twin specification. In: Annual Modeling
and Simulation Conference (ANNSIM). IEEE, pp 233–244
29. Barricelli BR, Casiraghi E, Fogli D (2019) A survey on digital twin: definitions, characteristics,
applications, and design implications. IEEE Access 7:167653–167671
30. David J et al (2020) Characterising the digital twin: a systematic literature review. CIRP J
Manuf Sci Technol 29:36–52
31. VanDerHorn E, Mahadevan S (2021) Digital twin: generalization, characterization and imple-
mentation. Decis Support Syst 145:113524
32. Mengnan L et al (2021) Review of digital twin about concepts, technologies, and industrial
applications. J Manuf Syst 58:346–361
33. Tao F et al (2022) Digital twin modeling. J Manuf Syst 64:372–389. ISSN: 0278-
6125. https://doi.org/10.1016/j.jmsy.2022.06.015. https://www.sciencedirect.com/science/
article/pii/S0278612522001108
34. Tao F et al (2021) Theory of digital twin modeling and its application. Comput Integr Manuf
Syst 27:1–15
35. Thelen A et al (2022) A comprehensive review of digital twin–part 1: modeling and twinning
enabling technologies. Struct Multidisc Optimization 65(12):354
36. Cronrath C, Ekström L, Lennartson B (2020) Formal properties of the digital twin—
implications for learning, optimization, and control. In: 2020 IEEE 16th international Con-
ference on Automation Science and Engineering (CASE). IEEE, pp 679–684
37. Zeigler BP (1976) The hierarchy of system specifications and the problem of structural infer-
ence. In: PSA: proceedings of the biennial meeting of the philosophy of science association,
vol 1976, 1. Cambridge University Press, pp 226–239
38. Worden K et al (2020) On digital twins, mirrors, and virtualizations: frameworks for model
verification and validation. In: ASCE-ASME J Risk Uncertainty Eng Syst Part B: Mech Eng
6(3):030902
39. Zheng X et al (2017) Perceptions on the state of the art in verification and validation in
cyber-physical systems. IEEE Syst J 11(4):2614–2627. https://doi.org/10.1109/JSYST.2015.
2496293
40. Khan A et al (2018) Digital Twin for legacy systems: simulation model testing and validation.
In: 2018 IEEE 14th international Conference on Automation Science and Engineering (CASE),
pp 421–426. https://doi.org/10.1109/COASE.2018.8560338.
41. Hasan A et al (2023) Predictive digital twins for autonomous surface vessels. In: Ocean
engineering, vol 288, p 116046. ISSN: 0029-8018. https://doi.org/10.1016/j.oceaneng.2023.
116046. https://www.sciencedirect.com/science/article/pii/S0029801823024307. A formal
framework for digital twin modeling, verification and validation 25
42. Argota Sánchez-Vaquerizo J (2022) Getting real: the challenge of building and validating
a large-scale digital twin of Barcelona & rsquo;s traffic with empirical data. ISPRS Int J
Geo-Information 11(1). ISSN: 2220-9964. https://doi.org/10.3390/ijgi11010024. https://www.
mdpi.com/2220-9964/11/1/24
142 M. K. Traore et al.

43. Zheng X, Julien C (2015) Verification and validation in cyber physical systems: research chal-
lenges and a way forward. In: 2015 IEEE/ACM 1st international workshop on software engi-
neering for smart cyber-physical systems, pp 15–18. https://doi.org/10.1109/SEsCPS.2015.
11
44. Wallace DR, Fujii RU (1989) Software verification and validation: an overview. IEEE Softw
6(3):10–17
45. Traore MK, Gorecki S, Ducq Y (2022) A simulation based approach to digital twin’s inter-
operability verification & validation. In: “Workshop interoperability challenges and solutions
within industrial networks” co-located with 11th international conference on Interoperability
for Enterprise Systems and Applications (I-ESA 2022), vol 3214
46. Gore R, Diallo S (2013) The need for usable formal methods in verification and validation. In:
2013 Winter Simulations Conference (WSC), pp 1257–1268. https://doi.org/10.1109/WSC.
2013.6721513
47. Collofello JS (1988) Introduction to software verification and validation. Carnegie Mellon
University, Software Engineering Institute
48. Fishwick PA, Miller JA (2004) Ontologies for modeling and simulation: issues and approaches.
In: Proceedings of the 2004 Winter Simulation Conference, vol 1. IEEE
49. Thacker BH et al (2004) Concepts of model verification and validation
50. Wang C et al (2018) Detecting aberrant behavior and item preknowledge: a comparison of
mixture modeling method and residual method. J Educ Behav Stat 43(4):469–501
51. Rasmus B et al (2008) Cross-validation of component models: a critical look at current methods.
Analy Bioanalytical Chem 390:1241–1251
52. Darlington RB, Hayes AF (2016) Regression analysis and linear models: concepts, applications,
and implementation. Guilford Publications
53. Yue J (2009) Variance analysis in software fault prediction models. In: 20th international
symposium on software reliability engineering. IEEE, pp 99–108
54. Koch K-R (2013) Parameter estimation and hypothesis testing in linear models. Springer Sci-
ence & Business Media
55. Kim WC, Ahn SC, Kwon WH (1995) Stability analysis and stabilization of fuzzy state space
models. Fuzzy Sets Syst 71(1):131–142. 26 Mamadou Kaba Traore Simon Gorecki and Yves
Ducq
56. Gidskehaug L, Anderssen E, Alsberg BK (2008) Cross model validation and optimisation of
bilinear regression models. Chemometrics Intell Lab Syst 93(1):1–10
57. Confalonieri R et al (2009) Multi-metric evaluation of the models WARM, CropSyst, and
WOFOST for rice. Ecol Modelling 220(11):1395–1410
58. Willmott CJ, Matsuura K (2005) Advantages of the mean absolute error (MAE) over the root
mean square error (RMSE) in assessing average model performance. Climate Res 30(1):79–82
59. Murat K, Arjan D (2017) Quality of service (QoS) in software defined networking (SDN): a
survey. J Netw Comput Appl 80:200–218
60. Data Warehousing Institute. Data quality and the bottom line: achieving business suc-
cess through a commitment to high quality data. http://www.dw-institute.com/ (visited on
02/08/2024)
61. Chengalur-Smith ISN, Ballou DP, Pazer HL (1999) The impact of data quality information on
decision making: an exploratory analysis. IEEE Trans Knowl Data Eng 11(6):853–864
62. Monica S (2006) Data quality: concepts. Data-Centric Systems and Applications, Springer,
Methodologies and Techniques
63. Office of Management and Budget. Information quality guidelines for ensuring and maximizing
the quality, objectivity, utility, and integrity of information disseminated by agencies. http://
www.whitehouse.gov/omb/fedreg/reproducible.html (visited on 01/17/2006)
64. DQI (2004) Data quality initiative framework. Project report. https://nces.ed.gov/fcsm/pdf/
FCSM.20.04_A_Framework_for_Data_Quality.pdf (visited on 2004)
65. Batini C et al (2009) Methodologies for data quality assessment and improvement. ACM
Comput Surv (CSUR) 41(3):1–52
A Formal Framework for Digital Twin Modeling, Verification, and Validation 143

66. Bertolazzi P, De Santis L, Scannapieco M (2003) Automatic record matching in cooperative

information systems. In: Proceedings of the international workshop on Data Quality in Coop-
erative Information Systsems (DQCIS), p 9
67. Hernández MA, Stolfo SJ (1998) Real-world data is dirty: data cleansing and the merge/purge
problem. Data Mining Knowl Discovery 2:9–37
68. Muthu S, Whitman L, Hossein Cheraghi S (1999) Business process reengineering: a consol-
idated methodology. In: Proceedings of the 4th annual international conference on industrial
engineering theory, applications and practice. University of Texas San Antonio TX, pp 17–20
69. Hammer M (1990) Reengineering work: don’t automate, obliterate. Harvard Bus Rev
68(4):104–112
70. Calinescu R et al (2015) Formal verification with confidence intervals to establish quality of
service properties of software systems. IEEE Trans Reliab 65(1):107–125
71. Tsai W-T, Chen Y, Paul R (2005) Specification-based verification and validation of web ser-
vices and service-oriented operating systems. In: 10th IEEE international workshop on object-
oriented real-time dependable systems. IEEE, pp 139–147
72. Bai X et al (2007) A framework for contract-based collaborative verification and validation
of web services. In: Component-based software engineering: 10th international symposium,
CBSE 2007, Medford, MA, USA, July 9–11. Proceedings 10. Springer, Heidelebrg, pp 258–273
73. Diakité M, Traoré MK (2024) Formalizing a framework of inference capabilities for Digital
Twin engineering. In: Simulation, p 00375497241228281
74. Dahmen U, Roßmann J (2018) Simulation-based verification with experimentable digital twins
in virtual testbeds. Tagungsband des 3. Kongresses Montage Handhabung Industrieroboter.
Springer, Heidelberg, pp 139–147
75. Clarke EM, Wing JM (1996) Formal methods: state of the art and future directions. ACM
Comput Surv (CSUR) 28(4):626–643
76. Cortes LA, Eles P, Peng Z (2003) Modeling and formal verification of embedded systems based
on a Petri net representation. J Syst Archit 49(12–15):571–598
77. Carroll J, Long D (1989) Theory of finite automata with an introduction to formal languages
78. Rushby JM (1995) Model checking and other ways of automating formal methods. In: Position
paper for panel on model checking for concurrent programs, Software Quality Week, San
Francisco
79. Goguen JA (2021) Theorem proving and algebra. In: arXiv preprint arXiv:2101.02690
80. Abarbanel Y et al (2000) Focs–automatic generation of simulation checkers from formal spec-
ifications. In: Computer aided verification: 12th international conference, CAV 2000, Chicago,
IL, USA, July 15–19, 2000. Proceedings 12. Springer, Heidelberg, pp 538–542
Digital Twins for Advanced
Manufacturing: The Standardized
Approach

Guodong Shao, Deogratias Kibira, and Simon Frechette

Abstract Digital Twins are becoming more prevalent in a wide range of industries
such as manufacturing, construction, smart city, and healthcare for various purposes,
including observing, predicting, optimizing, and controlling. Digital Twins are in
the early adoption stage. Currently, few standards directly address Digital Twins
and a commercial ecosystem of Digital Twins has not been well established. Devel-
oping and implementing Digital Twins present significant challenges. Most current
Digital Twin applications are customized solutions, which are expensive to create
and difficult to integrate with other systems. Foundational work is needed to support
an open marketplace for Digital Twin developers, users, and technology service
providers. This includes the development of standardized frameworks, reference
models, and interfaces to provide a solid foundation for ensuring interoperability,
reliability, validity, security, and trust. This Chapter identifies implementation chal-
lenges for Digital Twins for manufacturing, reviews relevant standardization efforts,
introduces the ISO Digital Twin framework standard for manufacturing, presents use
cases, and discusses potential research topics and future standardization directions.

Keywords Digital twin · Standards · Advanced manufacturing · Robot workcell

G. Shao (B) · D. Kibira · S. Frechette

Engineering Laboratory, National Institute of Standards and Technology, 100 Bureau Drive,
Gaithersburg, MD 20899, USA
e-mail: [email protected]
D. Kibira
e-mail: [email protected]
S. Frechette
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 145
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_7
146 G. Shao et al.

1 Introduction

The idea of developing physical mockups of planned objects, infrastructure, or scenes

has been practiced by humans for millennia. NASA provides one of the notable
recent applications of twinning during the planning and execution of space missions.
Mission planners created two similar space vehicles. One was sent on a mission while
its physical “twin” remained on Earth. The vehicle on Earth emulated the state of
the flying twin and could be used to test possible solutions to problems encountered
with the flying twin [1]. As computer technology advanced, physical models gave
way to digital models. The Digital Twin concept gained recognition in 2002 after
Dr. Grieves presented his vision of real space, virtual space, and information flowing
between real and virtual spaces [2]. The key idea is the state synchronization of the
real space and virtual space. However, it was not until recently that this concept
became one of the top strategic technology trends. The advancement of technologies
such as the Internet of Things (IoT), smart sensors, Artificial Intelligence (AI), and
cloud computing facilitate the realization of Digital Twins.
Digital Twins involve highly complex functional subsystems, including data
collection, data processing, communication, information modeling, data analytics,
visualization, simulation, optimization, and control. Some of these could be
distributed systems. This complex system of systems presents significant challenges
for manufacturers to understand and seamlessly integrate these diverse functional
subsystems. Many companies already have some forms of digital transformation
efforts underway. Others may already have multiple versions of digital solutions
from various vendors. Some vendors claim that they have complete solutions for
Digital Twin development. However, it is impossible to have all companies discard
existing digital systems and source new replacements from the same vendor because
not all manufacturers are technically able or can afford to redesign and replace their
entire plants. Therefore, existing systems and technologies must be integrated when
implementing new Digital Twins. Interoperability standards are needed to support
the communication and integration between (1) the physical and virtual systems, (2)
multiple Digital Twins, and (3) Digital Twins and legacy systems such as Manufac-
turing Execution Systems (MES), Enterprise Resource Planning (ERP), and Product
Life Cycle Management (PLM).
A Digital Twin comprises three main components—the physical system, the
virtual representation of the physical system, and bi-directional communication
between the physical and virtual worlds. Recent studies have introduced two more
dimensions: data and service [3]. It has then been noted that a framework is needed
to realize a Digital Twin, which comprises different components so developers can
use it for various applications and domains. While each Digital Twin may differ
in composition details, a high to medium-level general framework can help reduce
the Digital Twin development effort. Researchers have developed frameworks for
various aspects of Digital Twin development. Frameworks that emphasize the two-
way communication between the physical twin and the Digital Twin have been
Digital Twins for Advanced Manufacturing: The Standardized Approach 147

proposed [4, 5]. Many Digital Twins today are implemented for real-time state moni-
toring. The aerospace industry has used Digital Twins to monitor jet engines for many
years. These types of Digital Twins use one-way synchronization. They receive data
from the physical object but do not provide control feedback.
Galli et al. [6] analyzed a variety of proposed frameworks for building Digital
Twins. This research focused on the structure of the frameworks to find correlations
in terms of form and conceptualization. The results showed three types of archi-
tectures used for the Digital Twin. The first is the traditional and is based on the
original framework proposed by Grieves and Vickers [7]. This architecture supports
the Digital Twin in paralleling the real system, discussing interfaces and interoper-
ability with operations management systems. It also ensures synchronization between
the physical system and its corresponding digital twin. The second is the “service-
oriented” type, which consists of four components: a physical shop floor, a virtual
model of the shop floor, a service system, and the shop floor Digital Twin data. The
third architecture is the fractal, where local or specialized Digital Twins make up
the global Digital Twin, each with a similar structure. Khan et al. proposed a six-
dimensional framework to include (1) a physical asset, (2) a digital duplicate of the
physical asset, (3) data generated by the physical asset, (4) programs that improve
product performance and make the production process more efficient, (5) spiral-rings
like iterations that generate a more optimized product or efficient production process,
and (6) synchronization between the physical asset and the digital twin [4].
Despite these efforts, the Digital Twin technology is still in its early stages. There
is a lack of universal definitions, implementation frameworks, and protocols. There is
also a lack of comprehensive and in-depth analysis of Digital Twins from the perspec-
tive of concepts, technologies, and industrial applications research [8]. Further, the
ecosystem of Digital Twins is not well established, and most Digital Twins are devel-
oped using customized solutions. Customized solutions do not support reuse and are
costly and time-consuming. A systematic approach is needed to characterize and
manage Digital Twin subsystems to ensure cross-disciplinary interoperability and
credibility of the Digital Twin. Standards are the solutions to enable such interop-
erability. However, only a few standards have been developed for Digital Twins so
far.
According to Accenture research, companies are not taking full advantage of
Digital Twins because most Digital Twin applications are standalone for single func-
tions, which focus on functional optimization instead of enterprise optimization, and
have no comprehensive strategy for data integration and sharing [9]. Additionally, if
a Digital Twin is the current representation of a physical asset, the associated Digital
Twin data at any life cycle stage could be helpful for future asset management.
The architecture or mechanism supporting the flow of information about a product’s
performance and use from design, production, use, disposal, and recycling is referred
to as the digital thread [10]. Using the digital thread will enable the traceability of
Digital Twins from requirements to the retirement of the physical asset. Interoper-
ability among Digital Twins for different life cycle stages through digital threads
should help overcome these challenges. Standardized data representation will help
avoid customized Digital Twin development and the duplication of efforts. Digital
148 G. Shao et al.

threads can also provide an integrated view of the physical asset for Digital Twin
development, avoiding redundancy during information exchange.
To make the best use of Digital Twins, manufacturers need to apply interoperability
standards from both systems of systems and life cycle perspectives. In addition,
standards on vocabulary, reference architecture, and trustworthiness can help ensure
the interoperability, value, and credibility of the Digital Twins. These standards
will enable manufacturers to build, manage, and deploy their Digital Twins more
efficiently.
This Chapter focuses on a standardized approach to building Digital Twins for
advanced manufacturing. It identifies current challenges for manufacturers to imple-
ment their Digital Twins, reviews relevant standards efforts, introduces the ISO
Digital Twin framework standard, presents use-case scenarios, and discusses some
potential standards development research directions. The rest of the chapter is orga-
nized as follows: Sect. 2 discusses various applications of Digital Twins, Sect. 3
presents how standards help address the challenges of implementing and adopting
Digital Twins that manufacturers face, Sect. 4 introduces the ISO standard, ISO
23247—Digital Twin Framework for Manufacturing, Sect. 5 discusses additional
relevant standards, Sect. 6 describes a case study of applying standards for building
a Digital Twin of a robot work cell, Sect. 7 discusses potential topics and research
directions toward standardization, and Sect. 8 summarizes the chapter and presents
future work.

2 Applications and Enabling Technologies for Digital Twins

Digital Twins are becoming increasingly prevalent in a wide variety of industries,

including manufacturing, construction, smart cities, and healthcare. The applica-
tions of Digital Twins include system monitoring, anomaly detection, prediction,
optimization, and control. The major general classifications of Digital Twins, shown
in Fig. 1, are descriptive, diagnostic, predictive, prescriptive, and intelligent Digital
Twins. From left to right, Fig. 1 shows the different Digital Twin functionalities, with
increasing complexity from monitoring to intelligent control, to support appropriate
decision-making and automation. The bulletized items in each box are examples of
enabling technologies for that category of Digital Twins. Each category is described
in the following [11].
. Descriptive Digital Twins observe their physical counterparts to identify what has
happened or is happening. These kinds of Digital Twins can generate different
views of the data collected using smart sensors based on the purpose of the Digital
Twin. Based on stakeholders’ requirements, data and parameters can be visualized
in the form of text, tables, and charts. For example, one of these important data
is the cycle time of each product type during production. Enabling technologies
may include real-time data streaming, database queries, and dashboard reporting.
Digital Twins for Advanced Manufacturing: The Standardized Approach 149

Observe Analyze Predict Influence & Learn &

control Correct
Predictive
Descriptive Diagnostic Prescriptive Intelligent
Digital
Digital Twin Digital Twin Digital Twin Digital Twin
Twin

What happened What’s wrong How can we How can we

What is likely to
or is happening? and why is it make it happen? automatically
happen?
happening? achieve our
goals?
 Simulation
 Simulation
 Statistics & linear  Simulation
 Data stream  Data mining & regression  Simulation  Optimization
 Database forensics  Predictive data  Machine  Artificial
queries  Real-time mining learning Intelligence (AI)
 Reporting & analytics &  Forecasting &  Optimization  Automation
dashboards mining trend reporting  Control  Control

Increasing complexity, more decision support, and greater value

Fig. 1 Various applications of digital twins with examples of supporting technologies

. Diagnostic Digital Twins analyze what has gone wrong and why it has happened
or is happening to their physical counterparts. Diagnosis includes analyses of the
impact of a data input and an operational strategy on key performance indicators
(KPIs). For example, a product’s cycle time increase may be caused by machine
breakdowns or bad scheduling decisions. Diagnostics are supported by technolo-
gies and methodologies such as simulation, data mining, machine learning, and
data analytics.
. Predictive Digital Twins predict what will happen and when it will happen. These
kinds of Digital Twins can be used to estimate when a machine tool or robot
deterioration will likely reach a point of failure based on past performance patterns.
They can also pinpoint the cause or source of failure. Enabling technologies may
include modeling and simulation, predictive data mining, and parameter tracking.
For example, Digital Twins integrated with machine learning can be used to predict
cycle times for incoming manufacturing orders [5].
. Prescriptive Digital Twins provide the influence and control of the physical coun-
terparts and decide how to make it happen based on the objectives of the Digital
Twins. These kinds of Digital Twins can help identify the strategies and inputs
leading to optimal performance. For example, prescriptive analytics provide the
best possible input parameters and methods that enable cycle-time reduction and
increase throughput. Enabling technologies may include simulation, optimization,
and control.
. Intelligent Digital Twins are envisaged to control their physical counterparts based
on the strategies and parameters identified by the prescriptive Digital Twins. These
kinds of Digital Twins can learn new strategies based on the data collected and
take actions accordingly, which could include dynamically adjusting themselves
based on the changes in their physical counterparts to keep up with them or based
150 G. Shao et al.

on the Digital Twin objectives to ensure the physical counterparts operate opti-
mally. Artificial Intelligence (AI), including large language models, is the primary
modeling technique used for modeling the intelligent Digital Twin. Other enabling
technologies may include simulation, optimization, automation, and control.

The kind of Digital Twin to be implemented depends on the use case and is driven
by the objective and scope of the Digital Twin.

3 How Standards Support Digital Twin Development

and Integration

As discussed in Sect. 1, there are still significant challenges for manufacturers, espe-
cially Small and Medium-sized Enterprises (SMEs), to implement their Digital Twin
applications efficiently and effectively. Current implementations mostly use ad hoc,
customized approaches. Customized solutions not only increase the development
time and cost but also make it challenging to integrate with other systems and do not
support reuse. Standards are needed for manufacturers to go beyond custom, expen-
sive Digital Twins to an affordable marketplace of products and tools for Digital
Twins. Standards, such as frameworks, reference models, and interfaces, will provide
a solid foundation for Digital Twin developers, users, and technology and service
providers to ensure interoperability, reliability, validity, security, and trust.
Standards for Digital Twins will facilitate the composition and integration of
Digital Twins by providing guidelines, methodologies, common terminologies, archi-
tectures, and interface specifications. These standards can help make the creation,
integration, update, and validation of Digital Twins more accurate and consistent.
Standards can also help formalize requirements for Digital Twin projects, enable the
use of building blocks for Digital Twin implementations, analyze Digital Twin perfor-
mance, communicate between suppliers, partners, and customers, secure Digital
Twin information and protect privacy, and facilitate the verification and validation
of Digital Twins according to stakeholders’ requirements. Ultimately, standards will
help achieve “plug and play,” i.e., enabling interoperability between Digital Twins
and among software and hardware from various vendors. In different areas, standards
support the following:
. Defining a common language for data representation, communication protocols,
and Application Programming Interfaces (APIs) to ensure Digital Twin systems
can understand and interact with each other’s data. Examples of APIs include
those for data query, data update, and data synchronization.
. Developing shared metadata and ontologies to describe the properties, attributes,
and relationships within Digital Twins for easy mapping and translating data
between Digital Twins.
Digital Twins for Advanced Manufacturing: The Standardized Approach 151

. Developing tools or middleware that can map data from one Digital Twin’s format
to another using techniques such as data transformation, data normalization, or
data translation.
. Implementing robust security measures to ensure only authorized Digital Twins
can access and interact with each other using techniques such as authentication,
encryption, and access control.

4 Digital Twin Framework for Manufacturing

The ISO standard, ISO 23247—Digital Twin Framework for Manufacturing, was
created to facilitate the implementation of Digital Twins in manufacturing. The stan-
dard defines a “Digital Twin in Manufacturing” as a “fit for purpose digital repre-
sentation of an observable manufacturing element with synchronization between
the element and its digital representation [12].” It provides a generic guideline,
a reference architecture, and a framework for Digital Twin applications in manu-
facturing. The standard also provides examples of data collection, data commu-
nication, integration, modeling, and applications of relevant standards [12]. The
standard provides procedures for manufacturers and solution providers to analyze
Digital Twin requirements, define scope and objectives, use common terminologies,
comply with a generic reference architecture, and integrate multiple existing stan-
dards for various purposes. The framework includes the sub-entities and components
as building blocks for manufacturers to pick and choose for their own case-specific
Digital Twin development. It helps manufacturers systematically identify and deter-
mine subsystems and components, their relationships, and the characteristics of their
interactions from which appropriate standards can be selected for interoperability.
A key feature of ISO 23247 is that it enables the deployment of the digital thread,
implying that model-based engineering standards for various stages of a product life
cycle can be included in the framework. For example, for the Digital Twins devel-
oped to support a product at different stages, including design, manufacturing, and
inspection, relevant standards such as Standard for the Exchange of Product Model
Data (STEP) [13], MTConnect [14], and Quality Information Framework (QIF) [15]
can be applied. Therefore, the standard supports Digital Twins’ compatibility and
interoperability throughout the life cycle stages, allowing information reuse and
traceability.
The standard series includes four parts: (1) overview and general principles, (2)
reference architecture, (3) digital representation, and (4) information exchange. The
reference architecture in the standard includes a reference model with domains and
entities. There are four domains (layers), each with a logical set of tasks and functions
performed by functional entities. Figure 2 shows the entity-based reference model
and an illustration of the four domains and their interactions [16]. Each domain is
briefed as follows:
152 G. Shao et al.

Fig. 2 Functional view of the digital twin reference model for manufacturing

. The observable manufacturing domain: contains the Observable Manufacturing

Elements (OMEs), including any physical artifact, process, or behavior such as
personnel, equipment, material, processes, facilities, assets, and systems on the
factory floor. These OMEs can be represented by Digital Twins.
. The device communication domain: is a layer between OMEs and their Digital
Twins to support data exchange and synchronization. OMEs are monitored, and
real-time data are collected using device sensors and standard protocols in the
OME domain. This domain is also responsible for transferring commands and
signals for control and actuation of the OMEs.
. The Digital Twin domain, or core domain: is responsible for the modeling, oper-
ating, and managing Digital Twins. This domain hosts models, applications, and
services such as data analytics, simulation, and optimization to support provi-
sioning, monitoring, analysis, and synchronization. It also interacts with users
and systems, including other Digital Twins. As indicated in the definition, Digital
Twins are built “fit-for-purpose” because each has its own objectives, is context-
dependent, and only requires relevant data and models. The purpose of the
digital twin dictates its information content, model fidelity, and frequency of
synchronization.
. The user domain: includes users or systems such as a human, a device, an applica-
tion, or a system that uses applications and services provided by the Digital Twin
domain.

The cross-system entity in Fig. 2 is an entity that resides across domains to provide
common functionalities such as data translation, data assurance, and security support.
Digital Twins can be developed based on the Digital Twin Framework depicted within
the dotted line in Fig. 2. The framework supports the applications of IoT infrastructure
for data collection, communication protocols for data transmission, and information
flows between entities of different domains—OMEs, Data Collection and Device
Control, Digital Twin Core, and User layers.
Digital Twins for Advanced Manufacturing: The Standardized Approach 153

Developers tested the Standard by implementing several industry use cases to

demonstrate and validate the Standard. The use-cases included a Digital Twin for
robot drill and fill to increase equipment utilization, a Digital Twin that optimizes
the sizes of aircraft fasteners, and a Digital Twin to optimize Computer Numerical
Control (CNC) cutting tool life. These use-case implementations provided valuable
feedback to standards developers and demonstrated the viability of the standard
framework for constructing Digital Twins [17].

5 Additional Relevant Standards

Digital Twin standards, such as architectural frameworks, can provide implementa-

tion guidelines for developing Digital Twins. Other existing standards can support
specific Digital Twin functionalities such as data collection, data communication,
information modeling, systems integration, simulation modeling, and automation
and control. This section identifies some examples of manufacturing standards that
could apply to various aspects of the Digital Twin development. Standards related to
information security, data assurance, and trustworthiness are also necessary, but not
listed in this Section.

5.1 Frameworks and Architectures

. IEC 62832-1: 2020, Digital Factory Framework defines a framework to establish

and maintain the digital representations of production systems throughout their
life cycle. This framework supports a consistent exchange of information between
all processes and partners. Information becomes understandable, reusable, and
exchangeable throughout the production system life cycle [18].
. IEEE P2806:2019, System Architecture of Digital Representation for Physical
Objects in Factory Environments, supports the development of digital factories.
It describes the objective, components, data sources required, and procedure of
digital representation in factory environments [19].
. IEC 63278-1: Asset Administration Shell for industrial applications—Part 1:
Asset Administration Shell (AAS) structure [20]. AAS aims to enable one or
more software applications to exchange information and use that information in
a trusted and secure way. It specifies the connector between the real and virtual
worlds and includes a model of the shell covering the fundamental concepts: Asset,
Submodel, and Concept Description. Identifiers are defined for all elements in
the model, concept descriptions, and property definitions of external repositories
such as ECLASS and IEC CDD. Mappings of the AAS model are specified for
several widely used information models such as XML, JSON, RDF, OPC-UA,
and AutomationML.
154 G. Shao et al.

. ISO/IEC 30141:2018, Internet of Things (IoT)—Reference Architecture provides

a common vocabulary, reusable designs, and industry best practices. It starts with
collecting the essential characteristics of IoT, abstracting them into a generic IoT
conceptual model, and deriving a high-level systematic reference with subsequent
dissection of that model into five architectural views [21].
. ISO/IEC 21823-1:2019, Internet of things (IoT)—Interoperability for IoT
systems—Part 1: Framework provides an overview of the interoperability of IoT
systems and the various entities within them. It enables IoT systems to be built so
the entities of the IoT system can exchange information efficiently. It also supports
peer-to-peer interoperability between IoT systems [22].
. ISO/IEC/IEEE 15288:2015, Systems and software engineering—System life
cycle processes establish a common framework of process descriptions for
describing the life cycle of systems. It defines a set of processes and associ-
ated terminology that can be applied at any level in a system’s structure. They can
manage the stages of a system’s life cycle. It also provides processes that support
the definition, control, and improvement of the system life cycle processes [23].
. Microsoft Digital Twin Definition Language (DTDL)—A language for describing
models and interfaces for IoT Digital Twins. DTDL is based on JSON-LD and is
programming language independent. DTDL is used in different Microsoft services
such as IoT Hub, IoT Central, and Azure Digital Twins; it is also used to represent
device data in other IoT services such as IoT Plug and Play. DTDL covers the
resource description and not resource discovery and access. Resources (interfaces)
contain telemetry, properties, commands, relationships, and components [24].
. The High-Level Architecture (HLA) defines an architecture for distributed simu-
lation, its components, and the rules that outline the responsibilities of HLA feder-
ates and federations for a consistent implementation. The standard also supports
maintaining the information model of each simulation to retain its meaning
and purpose but enables data communication and time synchronization of the
distributed simulation systems [25].

5.2 Data Collection, Data Modeling, and Data Exchange

. MTConnect supports interoperability by providing a vocabulary for manufac-

turing equipment, making structured contextualized data possible, and avoiding
proprietary formats. Data sources of MTConnect include equipment, sensor
packages, and other factory floor hardware [14].
. OPC-Unified Architecture (UA) is a platform-independent standard used to
send messages between clients and servers over diverse networks with syntactic
interoperability [26].
. The MTConnect-OPC-UA Companion Specification supports interoperability
and consistency between MTConnect specifications and the OPC-UA spec-
ifications, as well as devices and software that implement those standards
[27].
Digital Twins for Advanced Manufacturing: The Standardized Approach 155

. ISO/IEC 20922, Information Technology—Message Queuing Telemetry Trans-

port (MQTT) v3.1.1, is a data protocol that supports client and server messaging
transport for publishing and subscribing. It is open and simple in design and
suitable for use in machine-to-machine (M2M) communication and IoT contexts
[28].
. ISO/IEC 17826, Information Technology—Cloud Data Management Interface
(CDMI), specifies how to access and manage stored cloud data [29].
. ISO 13374 series, Condition Monitoring and Diagnostics of Machines—Data
Processing, Communication, and Presentation, provides basic requirements for
open software specification [30].
. ISO/IEC 30161:2020, Internet of Things (IoT)—Requirements of IoT data
exchange platform for various IoT services, specifies requirements for an IoT data
exchange platform for: (1) middleware components of communication networks
allowing the co-existence of IoT services with legacy services, (2) end-points
performance across the communication networks among IoT and legacy services,
(3) IoT-specific functions allowing the efficient deployment of IoT services, (4)
IoT service communication networks’ framework and infrastructure, and (5) IoT
service implementation guideline for IoT data exchange platforms [31].
. Automation Markup Language (AutomationML) is a neutral data format based
on XML that can store and exchange information for plant engineering,
connecting heterogeneous modern engineering tools. Disciplines include mechan-
ical plant engineering, electrical design, human–machine interface development,
Programmable Logic Controller (PLC), and robot control [32].
. The Core Manufacturing Simulation Data (CMSD) standard provides an infor-
mation model schema to support data representation and exchange between
simulation and other manufacturing applications [33].
. ISO 15531, Industrial Automation Systems and Integration—Industrial Manu-
facturing Management Data supports information exchange between software
applications in production activities, including planning, scheduling, simulation,
control, and execution [34].
. ISO 14649-201; Industrial Automation Systems and Integration—Physical
Device Control—Data Model for Computer Numerical Control Controllers—
Part 201: Machine Tool Data for Cutting Processes describes technology specific
to machine tools. It defines data elements for manufacturing and machine
characteristics [35].
. ISO/IEC 21823-2:2020, Internet of things (IoT)—Interoperability for IoT
systems—Part 2: Transport interoperability, specifies a framework and require-
ments for transport interoperability to enable the construction of IoT systems
with information exchange, peer-to-peer connectivity, and seamless communi-
cation between different IoT systems and among entities within an IoT system
[36].
. ISO/IEC 21823-3:2021, Internet of Things (IoT)—Interoperability for IoT
systems—Part 3: Semantic interoperability, provides: (1) basic concepts for IoT
systems including requirements of the core ontologies for semantic interoper-
ability, (2) best practices and guidance on how to use ontologies to develop
156 G. Shao et al.

domain-specific applications, (3) relevant IoT ontologies along with compara-

tive study of the characteristics and approaches in terms of modularity, extensi-
bility, reusability, scalability, interoperability with upper ontologies, and (4) use
cases and service scenarios that exhibit necessities and requirements of semantic
interoperability [37].
. ISO/IEC 21823-4:2022, Internet of things (IoT)—Interoperability for IoT
systems—Part 4: Syntactic interoperability, describes five facets of IoT interop-
erability: transport, semantic, syntactic, behavioral, and policy. It includes spec-
ifications on how to achieve syntactic interoperability among IoT devices and a
framework for processes for developing information exchange rules related to IoT
devices [38].

5.3 Digital Representation

. The American Society of Mechanical Engineers (ASME) Y14.26M, Digital

Representation for Communication of Product Definition Data, focuses on the
representation and communication of data to define products. It supports product
data exchange developed in computer-aided manufacturing (CAM) systems [39].
. ISO 10303, Automation Systems and Integration—Product Data Representation
and Exchange, also known as Standard for the Exchange of Product model data
(STEP), supports the exchange of product manufacturing information [13].
. ISO 10303-IR 105, Automation Systems and Integration—Product Data Repre-
sentation and Exchange—Part 105: Integrated Application Resource: Kinematics
focuses on the representation of kinematics information of a mechanical product
[40].
. Quality Information Framework (QIF)—This framework standard enables the
capture, use, and reuse of metrology-related information throughout the Product
Life Cycle Management (PLM) and Product Data Management (PDM) domains.
It supports the creation of digital threads. It applies to product design, manufac-
turing, and quality inspection. It relies on the Extensible Markup Language (XML)
standard and contains a library of XML schemas. It supports data integrity and
interoperability in implementing model-based enterprise and IoT [15].
. ASME B5.59-2, Information Technology for Machine Tools Part 2, defines the
properties needed to describe machine tools used for milling and turning [41].
. ISO 13399, Cutting Tool Data Representation and Exchange, provides a model
and a reference dictionary to represent cutting tools. EXPRESS schema is used
for the product description, and product files can be generated according to the
schema [42].
. ISO 16400, Equipment Behaviors Catalogue (EBC) defines a template and rules
for describing equipment behaviors, such as state transition and time series of
operation results, that are produced because of machine activities to be registered
in the common repository. It specifies the methodology to construct catalogs of
equipment behavior to plan and analyze production system performance [43].
Digital Twins for Advanced Manufacturing: The Standardized Approach 157

. Predictive Model Markup Language (PMML) is used to develop predictive and

descriptive models and to represent pre- and post-processed data. PMML is based
on XML and supports the representation of statistical and data-mining models
and the model sharing between PMML-compliant applications. Examples of
models include neural networks, decision trees, Gaussian progress, and Bayesian
networks [44].

With these relevant standards from various functional categories, users can select
those applicable to their Digital Twin implementations.

6 A Use Case of Developing Digital Twins

with a Standardized Approach

In this section, a robotic work cell Digital Twin is discussed to exemplify the appli-
cation of relevant standards. The work cell consists of collaborative robot arms for
material handling and machine tending, a CNC machine tool for machining, and
a Coordinate Measuring Machine (CMM) for product geometry measurements and
quality control. Figure 3 shows the workflow and equipment for the use-case scenario
in the work cell. The workflow includes receiving parts, loading parts to the CNC
(by ROBOT #1), cutting the parts, unloading parts from the CNC (by ROBOT #1),
loading parts to the CMM (by ROBOT #2), inspecting the parts, and offloading
parts from the CMM (by ROBOT #2). Parts that fail the inspection will be sent to
the rework buffer by ROBOT#2. The cell has a single input location and a single
output location. The robots are fitted with a 2F-85 gripper to handle components
with different geometries effectively.
One of the research objectives is to build a process digital twin that is a composite
of the Digital Twins of each individual component.

Fig. 3 Workflow through machines and equipment in the robot work cell
158 G. Shao et al.

The standards used in implementing the use-case include ISO 23247, MTCon-
nect, STEP, QIF, and ASME Verification, Validation, and Uncertainty Quantification
(VVUQ).

6.1 ISO 23247

Based on the Digital Twin framework introduced in Sect. 3, each piece of equipment
in the work cell is treated as an OME, for which data need to be collected, and Digital
Twins for various scenarios must be developed. In this section, we focus on one robot
arm in the work cell to showcase the method of Digital Twin development.
The ISO Digital Twin framework standard is instantiated for the robot arm (UR5e),
whose operations include picking and loading a workpiece to a CNC machine tool.
Figure 4 illustrates the instantiation of the framework for the robot arm. The figure
shows data being collected from the robot arm through a MTConnect adapter. The
Digital Twin entity comprises the simulation model of the robot arm and analytical
models that manipulate real-time data to support decision-making. The modeling
method and environment support the three-dimensional geometry of the robot arm
components, including the robot base, links, joints, end effector, and workpiece. The
user entity includes the developer and user of the Digital Twin, production software
systems, or other Digital Twins. Similarly, the standard can be instantiated for the
cutting part, the CNC machine, and the CMM.

6.2 MTConnect

The process of acquiring data and building a scalable data pipeline for the work cell
involves multiple activities: (1) collecting real-time operational data from the robot,
(2) leveraging the MTConnect standard to provide defined machine data and make it
available in a standard format, and (3) developing a set of tools to enable client-side
use of the MTConnect agent.
Operational Data Collection: To collect the operational data from the UR5e robot
arm, we utilized the Universal Robot’s Real-Time Data Exchange (UR-RTDE) inter-
face with APIs supporting data collection. Several data items collected include
angular position, velocity, acceleration, torque, current, and temperature for each
of the six joints of the robot arm. Universal Robots also shows the data items (and
the corresponding units) that the vendor offers through the UR-RTDE interface.
MTConnect Interface: Implementing the MTConnect standard requires an adapter
and an agent. The adapter serves as a data collection element from the equipment
while the agent collects data from the adapter. Many machine vendors provide a
preinstalled adapter but not the UR5e robot. The adapter was installed on an interface
connected to the robot arm. The adapter packages data into a format that is readable
Digital Twins for Advanced Manufacturing: The Standardized Approach 159

Fig. 4 Implementation method of building a Digital Twin for the robot based on ISO 23247

by the agent. The agent provides an application interface to retrieve the MTConnect
data gathered from the adapter. Figure 5 shows the data flow from a physical device
(UR5e) to the Digital Twin. A semantic structure was provided for the physical
data generated by the UR5e robot arm through MTConnect. This semantic structure
includes data tags and units based on the MTConnect 2.0 Standard [42]. Using the
Python UR-RTDE API, a socket-based adapter that sends MTConnect-compliant
data to the agent was developed. An instance of the MTConnect agent receives and
serves the data in a machine-readable format.
Client-Side Integration: The MTConnect agent is implemented in C++ and displays
data in XML format on a Hypertext Transfer Protocol (HTTP) server. The tool
used for modeling the robot work cell is the Simulink/Simscape, which can input
comma-separated values (CSV) where multiple data items over a time interval are

Fig. 5 Data pipeline for the UR5e robot Digital Twin

160 G. Shao et al.

Fig. 6 The physical and virtual model of the UR5e robot arm with attached gripper

recorded and synced up to their respective timestamps (UTC format). The tools have
been developed to parse the XML output from the MTConnect Agent, populate a
2-dimensional array, and store the array in a CSV file.

6.3 Step

To develop a physical model of the robot arm, Computer Aided Design (CAD) models
of its components and those of the work environment are fundamental. These models
include the description of the geometry of the links and how they are connected to
the robot arm. These CAD models are imported into the Digital Twin environment to
create a physical model of the UR5e robot arm. The CAD models for the gripper are
“assembled” into the end effector in the AutoCAD Inventor environment, exported
to the Digital Twin environment, and “attached” to the robot arm model. Figure 6
shows the physical and virtual models of the robot arm with the attached end effector.
The CAD models of the part, CNC, and CMM can be in STEP standard format,
which allows easy exchange and visualization. The design part model in STEP will
also be used to compare the measurements of the finished product.

6.4 Quality Information Framework (QIF)

QIF provides an integrated model for manufacturing quality information. The CMM
provides measurement information about a product regarding conformance to spec-
ifications per Geometric Dimensioning and Tolerancing (GD&T) design. Design
tolerances are defined by the amount a feature is allowed to vary from the nominal.
Digital Twins for Advanced Manufacturing: The Standardized Approach 161

Assigning GD&T to a designed part is a way to consider process variation within

manufacturing. This product data is obtained directly from the design process and
should be available throughout the product life cycle. Being able to map this data
back to a single source—the native CAD, QIF enables model-based workflows that
are part of the digital manufacturing transformation. The Digital Twin of the CMM
will receive measurement data from the actual machine and compare it with the
design data. QIF facilitates this data exchange between the real CMM and its Digital
Twin.

6.5 Verification and Validation

Digital Twins are complex systems that are sometimes used as virtual testbeds for
verification and validation of systems, especially in cases where actual tests are
complex to perform. However, the quality of decisions made with Digital Twins
depends on the validity of the underlying models. A valid Digital Twin should accu-
rately describe the system that changes over time. Thus, the development of the
Digital Twins needs to be validated before its use in supporting decision-making.
Verification, Validation, and Uncertainty Quantification (VVUQ) standards need to
be followed to ensure that the Digital Twin is built correctly and that the right Digital
Twin has been built. Zhang et al. [45] discuss verification and validation methods
for Digital Twins, which are categorized into qualitative, quantitative, and integrated
methods. Both qualitative and quantitative methods require metrics for Digital Twin
validation. These metrics include credibility/fidelity, complexity, standardization,
and capability maturity of model construction.
Hua et al. [46] summarized general strategies to validate a Digital Twin. These
include visual inspection of the Digital Twin for correctness using established stan-
dards, testing properties of the Digital Twin, model-based testing using methods such
as input–output conformance testing, and machine learning or artificial intelligence-
based testing. Kibira and Weiss [47] used a model-based approach to validate the
Digital Twin model of a robot arm. Joint position and orientation data, velocity data,
and acceleration data were collected from the physical twin for validation under the
model-based approach.
Verification and validation standards include:
. ASME V&V 10: Standard for Verification and Validation in Computational Solid
Mechanics [48].
. ASME V&V 20: Standard for Verification and Validation in Computational Fluid
Dynamics and Heat Transfer [49].
. ASME V&V 40: Assessing Credibility of Computational Modeling Through
Verification and Validation: Application to Medical Devices [50].
. Other ASME V&V 50, 60, 70, and 80 standards are under development. V&V
50 standards are for advanced manufacturing, and V&V 70 standards are for
data-driven models.
162 G. Shao et al.

7 Future Research Topics for Standardization

The current four parts of the ISO 23247 series provide a fundamental generic Digital
Twin framework for manufacturing. The framework can be extended to industries
that employ specialized manufacturing processes and technologies. Future work on
this standard may result in new additions supporting the development and validation
of Digital Twins. The new research topics for extension of the standard include (1)
digital thread for Digital Twins, (2) Digital Twin composition, (3) ontologies of the
Digital Twin framework to clarify the entities and relationships, (4) building Digital
Twins from reusable components to increase the consistency and reduce the develop-
ment time, (5) credibility assessment of Digital Twins to increase the trustworthiness
and value for decision-making, (6) Digital Twins and the metaverse to provide guide-
lines that enable the integration between Digital Twins and industrial metaverse, (7)
plug and play Digital Twin integration by standardizing interfaces with customers’
environment and application platforms, and (8) extending the framework to specific
sectors, e.g., semiconductor manufacturing, biomanufacturing, and additive manu-
facturing to address domain-specific needs [51]. The following subsections discuss
the potential new parts of the standard that could enable better and easier Digital
Twin development.

7.1 Digital Thread for Digital Twin (ISO 23247-5)

When performing digital transformation, standalone Digital Twins require a lot of

duplicated efforts. Therefore, a life cycle approach needs to be taken. However, effec-
tively bringing the life cycle approach is challenging. Guidelines and methodologies
for supporting Digital Twin development using a digital thread of the product life
cycle will be needed to access all product life cycle information, including design,
manufacturing, inspection, and use data, and enable information traceability.
A new part of ISO 23247 on this topic will specify how the digital thread enables
the creation, connectivity, management, and maintenance of Digital Twins across
the product life cycle. This will involve defining the principles to follow before
embarking on a journey for digital transformation, describing methodologies, and
providing use-case examples. This part will describe how the digital thread supports
the generation, implementation, and transformation of Digital Twins in manufac-
turing. Information in the digital thread enables the Digital Twin to express the
changes in a product throughout its life cycle, which can be used to improve future
iterations of the product. The digital thread ensures this product life cycle information
is readily accessible, traceable, reliable, and secure.
Digital Twins for Advanced Manufacturing: The Standardized Approach 163

7.2 Digital Twin Composition (ISO 23247-6)

Digital Twin composition implies that multiple Digital Twins are developed and
integrated with the support of a digital thread. For example, the Digital Twin of a
part and that of the machine that manufactures the part can interact dynamically
and seamlessly. When the Digital Twin of a cutting tool, a machine tool, and a part
interact, they can be used to determine the tool wear, tolerance conformance, and the
machine’s health. Digital Twins of multiple partners coordinate and communicate in
real time in a supply chain. However, it is challenging to aggregate, compose, and
integrate multiple Digital Twins and applications to achieve a new goal. Standard-
based methods and guidelines will help achieve this, reduce development time, and
mitigate risks for such undertakings.
This part of ISO 23247 will provide guidelines on enabling multiple Digital Twins
to communicate and interoperate effectively. The new part could provide generic
methodologies, principles, and examples to help users understand the purpose of
the Digital Twin and develop the appropriate Digital Twin to address the identified
problem(s). Relevant standards and technologies could be selected and applied to
demonstrate the integration.
This part of ISO 23247 will also specify Digital Twin composition by defining
principles, showing methodologies, and providing use-case examples of Digital
Twin configuration, communication, aggregation, composition, integration, and
collaboration during manufacturing.

7.3 Ontologies for Digital Twin Framework

The current four parts of the ISO 23247 standard define the terms, relationships,
components, and processes necessary for developing a Digital Twin and provide
guidelines for Digital Twin implementation. However, logical formalism does not
support it, which may lead to inconsistent implementation. The ontologies for the
Digital Twin framework will enable the definitions of terms that are both human-
understandable and computer-processable, which result in an unambiguous repre-
sentation of a particular construct and consistent interpretation, regardless of the
initial data source. It also enables explicit representation of the connections between
different terms; different connections permit a consistent presence and representa-
tion of the required metadata. A potential new part of ISO 23247 on this topic could
provide an ontology for the Digital Twin framework.
164 G. Shao et al.

7.4 Building Digital Twins from Reusable Components

Digital Twins could be developed for different control levels depending on the
application, including equipment, work cells, production lines, factories, and supply
chains. While some approaches exist to support model component reuse [52], most
of them are not explicitly designed for Digital Twins. Therefore, almost all Digital
Twins are built from scratch, which makes implementations time-consuming and
costly. Customized designs also make a Digital Twin challenging to modify, extend,
and reuse. Manufacturing knowledge, information attributes, and use-case configu-
rations are often developed using different specialized abstractions for each appli-
cation. The reusability of Digital Twin components in a Digital Twin library could
considerably reduce the development cost, time, and the required level of expertise.
A potential new part of ISO 23247 on this topic could provide guidelines on
building component libraries and creating templates for organizing data, informa-
tion, and models. Reusable Digital Twin components may include templates for
data collection, common information attributes, and modular models. Digital Twin
development would be supported by enabling technologies and relevant standards for
various Digital Twin functions. The new part could provide generic methodologies,
architectures, frameworks, knowledge bases, and examples for building and using
Digital Twin component libraries.

7.5 Credibility Assessment of Digital Twins

The current four parts of ISO 23247 do not cover VVUQ and testing on Digital Twins.
Given the potential use of Digital Twins in critical decision-making for various manu-
facturing applications, the results generated by Digital Twins must be trustworthy
for real manufacturing needs. Model credibility assessment, including VVUQ tech-
niques, must be applied throughout the life cycle of Digital Twins. VVUQ should be
embedded in the design, creation, and deployment of Digital Twins to establish trust
in the model and its outputs [53]. Verification and Validation (V&V) activities are
necessary to ensure that a Digital Twin meets its intended purpose and design goals.
Uncertainty Quantification (UQ) produces a measure of performance that users can
apply as part of a credibility assessment for a given Digital Twin. VVUQ for Digital
Twins should be a continual process that adapts to changes in the OME and its digital
representation, data inputs, and decisions made [53]. The credibility assessment of
Digital Twins may also include factors beyond VVUQ.
Digital Twin testing needs a test system comprising a set of tests for both the
OME and its Digital Twin. The test system should also define what an acceptably
valid Digital Twin should look like. Grieves proposed a virtual testing method for
manufactured products, which can be adapted to Digital Twins [54]. For example,
suppose the test system can run a set of tests, and the results of the Digital Twin can’t
be distinguished from those of the OME within a predefined probability threshold. In
Digital Twins for Advanced Manufacturing: The Standardized Approach 165

that case, the Digital Twin can be regarded as a reasonable representation of the OME.
Trust in a Digital Twin also involves trust in the data collected from the OME, the
model used in the Digital Twin, the data updating procedure, and the recommended
decisions. All these aspects should have a measurable uncertainty, whose existence
means that validation (comparison with reality) needs to be treated as a statistical
process. Comparison of actual data with model results can be used to estimate the
probability that the Digital Twin is a consistent representation of the OME.
Currently, there is no standard process for reporting VVUQ for digital twins.
Developing robust VVUQ processes for digital twins remains a challenge. A potential
new part of ISO 23247 on this topic could provide guidelines on and methodologies
for how to measure uncertainty, how to perform VVUQ and testing for Digital Twins,
how to select or construct a credibility assessment framework that supports these
activities, and how to assess the credibility of the developed Digital Twins.

7.6 Digital Twins and the Metaverse

The metaverse can support monitoring the manufacturing system in real time, both
visually and from a metric standpoint. It could also provide users with an immersive
experience. This is now possible because of the maturity of technologies for virtual
reality (VR), augmented reality (AR), and extended reality (XR), which can enhance
users’ visualization experience for manufacturing. For example, it has been demon-
strated that AR technologies can be integrated with three-dimensional geometrical
product specification and verification standards and practices [55].
A significant feature of the metaverse is the immersive visualization experience
along with its human–machine interface. The hardware and software technologies
developed for the metaverse can be used by the Digital Twin framework for manufac-
turing, especially in cases where there is human involvement. For example, the user
domain (shown in Fig. 2) and user entity of the ISO 23247 standard can use human–
machine interfaces provided by the metaverse. Alternatively, a metaverse may be a
parallel virtual world that may subsume some of the Digital Twins of a manufac-
turing enterprise that the metaverse represents. A new part of the ISO 23247 series
could include the metaverse concept, its definition, possible scenarios for integrating
with manufacturing Digital Twins, including human Digital Twins, guidelines, and
methodologies for such integration.

7.7 Extending the Framework to Specific Sectors

Based on the generic framework provided by the initial four parts of the ISO 23247
series, extensions can be developed as new parts of the standard for specific manufac-
turing sectors such as biomanufacturing, semiconductor manufacturing, and additive
166 G. Shao et al.

manufacturing. The new parts may include specialization of the Digital Twin frame-
work by adding new functional entities or modifying existing functional entities to
fit the new requirements. The new parts may also have the use cases implemented
for those manufacturing sectors. These use-case implementations may, in turn, help
identify new standardization requirements for that manufacturing sector.
The emerging biomanufacturing sector can use the generic framework to develop
its Digital Twins, which may constitute a new part of the ISO 23247 series. Similarly,
Digital Twins for additive manufacturing may have substantial potential to improve
process control, and a new part in the ISO 23247 series can be dedicated to additive
manufacturing. In semiconductor manufacturing, an extension of the standard can be
developed to address challenges such as obtaining datasets for constructing Digital
Twin models and cybersecurity associated with the Digital Twin. Other standards
development organizations may adopt the current ISO 23247 series to create Digital
Twins for their customer industries.

8 Summary

Digital Twins are becoming more prevalent in a wide variety of industries, including
manufacturing. However, a standards-based ecosystem of Digital Twins has not yet
been established. Developing and integrating Digital Twins presents significant chal-
lenges. Foundational work is needed to support an open marketplace for Digital Twin
developers, users, and technology and service providers. This includes the devel-
opment of standardized frameworks, reference models, interface specifications, and
VVUQ methodologies to provide a solid foundation for ensuring the interoperability,
validity, security, and trust of Digital Twins. This Chapter focuses on applying Digital
Twins in manufacturing within a framework of standards, identifies current chal-
lenges, reviews relevant standardization efforts, and introduces the ISO Digital Twin
framework standard for manufacturing, ISO 23247. This work also discusses some
potential research directions and future standardized topics. A use case is presented
to illustrate the Digital Twin development process by applying relevant standards.
The example is obtained from the Digital Twin development efforts for a robot work
cell in a Digital Twin Lab at NIST.
Future efforts include performing measurement science research to support the
development and integration of Digital Twins in manufacturing, working with Indus-
trial consortia and standards development organizations to prioritize the standard-
ization topics (e.g., interoperability and VVUQ), formulating working groups and
project teams to develop the new parts of the Digital Twin standards, and enhancing
the Digital Twin Lab to serve as a digital twin testbed to support digital twin
prototyping and standards development and testing.
Digital Twins for Advanced Manufacturing: The Standardized Approach 167

Disclaimer Certain commercial products and systems are identified in this chapter to facilitate
understanding. Such identification does not imply that these software systems are necessarily the
best available for the purpose. No approval or endorsement of any commercial product by NIST is
intended or implied.

References

1. Allen D (2021) Digital twins and living models at NASA [Slides]. In: ASME digital twin
summit. https://ntrs.nasa.gov/api/citations/20210023699/downloads/ASME%20Digital%20T
win%20Summit%20Keynote_final.pdf
2. Grieves M (2002) Completing the cycle: using PLM information in the sales and service
functions [Slides]. In: SME management forum. Troy, MI
3. Khan A, Shahid F, Maple C, Ahmad A, Jeon G (2022) Toward smart manufacturing using
spiral digital twin framework and twin chain. IEEE Trans Industr Inf 18(2):1359–1366
4. Negri E, Berardi S, Fumagalli L, Macchi M (2020) MES-integrated digital twin frameworks.
J Manuf Syst 56:58–71
5. Jain S, Narayanan A (2023) Digital twin–enabled machine learning for smart manufacturing.
Smart Sustain Manuf Syst 7(1):111–128
6. Galli E, Fani V, Bandinelli R, Lacroix S, Le Duigou J, Eynard B, Godart X (2023) Literature
review and comparison of Digital Twin frameworks in manufacturing. Proc Eur Counc Model
Simul 2023:428–434
7. Grieves M, Vickers J (2017) Digital Twin: Mitigating unpredictable, undesirable emergent
behavior in complex systems. In: Kahlen F-J, Flumerfelt S, Alves A (A c. Di) Transdisciplinary
perspectives on complex systems. Springer International Publishing, pp 85–113. https://doi.
org/10.1007/978-3-319-38756-7_4
8. Lu Y, Liu C, Kevin I, Wang K, Huang H, Xu X (2020) Digital Twin-driven smart manufacturing:
connotation, reference model, applications and research issues. Robot Comput-Integr Manuf
61:101837
9. Accenture (2021) Think thread first: surf the wave of product data. Retrieved 6 Oct 2023, from
https://www.accenture.com/us-en/insights/industry-x/thread-first-thinking
10. Singh V, Willcox KE (2018) Engineering design with digital thread. AIAA J 56(11):4515–4528
11. Shao G, Jain S, Shin J (2014) Data analytics using simulation for smart manufacturing. In:
Tolk A, Diallo SY, Ryzhov IO, Yilmaz L, Buckley S, Miller JA (eds) Proceeding of the 2014
winter simulation conference. IEEE, pp 2192–2203
12. ISO-23247-1 (2021) Automation systems and integration—Digital Twin framework for manu-
facturing—Part 1: overview and general principles. Retrieved on 10 Oct 2023, from https://
www.iso.org/standard/75066.html
13. ISO-10303 (2021) Industrial automation systems and integration—product data representation
and exchange—Part 1: Overview and fundamental principles, Retrieved on 21 Oct 2023, from
https://www.iso.org/standard/72237.html
14. MTConnect (2022) MTConnect standardizes factory device data. https://www.mtconnect.org/
15. ISO (2020) ISO 23952-2020. Automation systems and integration—Quality information frame-
work (QIF)—An integrated model for manufacturing quality information. Retrieved on 11 Oct
2023, from https://www.iso.org/standard/77461.html
16. ISO-23247-2 (2021) Automation systems and integration—Digital twin framework for manu-
facturing —Part 2: reference architecture. Retrieved on 10 Oct 2023, from https://www.iso.
org/obp/ui/en/#iso:std:iso:23247:-2:ed-1:v1:en
17. STEP Tools (2020) Demonstration of Three ISO 23247. Retrieved on 30 Nov 2023 from https://
www.youtube.com/watch?v=wbsC_qzB8us
168 G. Shao et al.

18. IEC-62832 (2020) Industrial process measurement, control and automation—digital factory
framework—Part1: General principles, Retrieved on 12 Oct 2023, from https://webstore.iec.
ch/publication/65858
19. IEEE-P2806 (2019) System architecture of digital representation for physical objects in factory
environments, Retrieved on 15 Oct 2023, from https://standards.ieee.org/project/2806.html
20. IEC 63278 -1. (2023) AAS (Asset Administration Shell) for industrial applications - Part 1:
Asset Administration Shell structure. https://webstore.iec.ch/en/publication/65628
21. ISO/IEC-30141 (2018) Internet of Things (IoT)—reference architecture, Retrieved on 11 Oct
2023, from https://www.iso.org/standard/65695.html
22. ISO/IEC-21823-1 (2019) Internet of things (IoT)—interoperability for IoT systems—Part 1:
framework, Retrieved on 11 Oct 2023, from https://www.iso.org/standard/71885.html
23. ISO/IEC/IEEE-15288 (2015)Systems and software engineering—system life cycle processes,
Retrieved on 24 Oct 2023, from https://www.iso.org/standard/63711.html
24. DTDL(Digital Twin Definition Language) (2022) Digital twin definition language (DTDL)
for models. Retrieved on 29 Sept 2023, from https://learn.microsoft.com/en-us/azure/digital-
twins/concepts-models#digital-twin-definition-language-dtdl-for-models
25. HLA (High Level Architecture) (2010) IEEE standard for modeling and simulation (M&S)
High level architecture (HLA)—federate interface specification. IEEE Comput Soc
26. OPC-UA (2017) OPC-UA services specification, Retrieved on 11 Oct 2023 from https://
opcfoundation.org/developer-tools/specifications-unified-architecture#:~:text=OPC%20Unif
ied%20Architecture%20Specification,more%20secure%20and%20scalable%20solution
27. MTConnect-OPC UA (2019) OPC-UA companion specification for MT connect. https://www.
mtconnect.org/opc-ua-companion-specification
28. ISO/IEC-20922 (2016) Information technology—Message queuing telemetry transport
(MQTT) v3.1.1, Retrieved on 26 Oct 2023, from https://www.iso.org/standard/69466.html
29. ISO/IEC-17826 (2016) Information technology—cloud data management interface (CDMI).
Retrieved on 26 Oct 2023, from https://www.iso.org/standard/70226.html
30. ISO-13374 (2007) Condition monitoring and diagnostics of machines—Data processing,
communication and presentation—Part 2: Data processing, Retrieved on 21 Oct 2023. From
https://www.iso.org/standard/36645.html
31. ISO/IEC-30161 (2022) Internet of Things (IoT)—requirements of IoT data exchange platform
for various IoT services, Retrieved on 23 Oct 2023, from https://www.iso.org/standard/53281.
html
32. Automation ML (2021) Standard data exchange in the engineering process of production
systems, Retrieved on 10 Oct 2023, from https://www.automationml.org/wp-content/uploads/
2021/06/AutomationML-Brochure.pdf
33. CMSD (2010) Core manufacturing simulation data (CMSD) standard, https://www.sisostds.
org/DesktopModules/Bring2mind/DMX/API/Entries/Download?Command=Core_Down
load&EntryId=31457&PortalId=0&TabId=105
34. ISO-15531 (2017) Industrial automation systems and integration—industrial manufacturing
management data—Part 44: Information modelling for shop floor data acquisition, Retrieved
on 26 Oct 2023, from https://www.iso.org/standard/71064.html
35. ISO-14649 (2011) Industrial automation systems and integration—physical device control—
data model for computerized numerical controllers—Part 201: Machine tool data for cutting
processes, Retrieved on 26 Oct 2023, from https://www.iso.org/standard/60042.html
36. ISO/IEC-21823-2 (2020) Internet of things (IoT)—interoperability for IoT systems—Part
2: Transport interoperability, Retrieved on 11 Oct 2023, from https://www.iso.org/standard/
80986.html
37. ISO/IEC-21823-3 (2021) Internet of things (IoT) - interoperability for IoT systems—Part 3:
semantic interoperability, https://www.iso.org/standard/83752.html
38. ISO/IEC-21823-4 (2022) Internet of things (IoT)—interoperability for IoT systems—Part 4:
Syntactic interoperability, Retrieved on 11 Oct 2023, from https://www.iso.org/standard/84773.
html
Digital Twins for Advanced Manufacturing: The Standardized Approach 169

39. ASME (American Society of Mechanical Engineers) (1989) Digital representation for commu-
nication of product definition data. Retrieved on 10 Oct 2023, from https://standards.globalspec.
com/std/437642/ASME%20Y14.26M
40. ISO-10303-105 (2014) Industrial automation systems and integration—product data represen-
tation and exchange—Part 105: Integrated application resource: Kinematics, Retrieved on 21
Oct 2023, from https://www.iso.org/standard/64294.html
41. ASME (American Society of Mechanical Engineers) (2009) Data specification for properties
of machining and turning centers, ASME B5.59-2
42. ISO-13399 (2006) Cutting tool data representation and exchange—Part 1: overview, funda-
mental principles and general information model, Retrieved on 26 Oct 2023, from https://www.
iso.org/standard/36757.html
43. ISO-16400 (2020) Automation systems and integration—equipment behavior catalogues for
virtual production system—Part 1: overview, Retrieved on 26 Oct 2023, from https://www.iso.
org/standard/73384.html
44. PMML (2018) The predictive model markup language (PMML) 4.3, Retrieved on 11 Oct 2023
from http://dmg.org/pmml/v4-3/GeneralStructure.html
45. Zhang L, Zhou L, Horn BK (2021) Building a right Digital Twin with model engineering. J
Manuf Syst 59:151–164
46. Hua EY, Lazarova-Molnar S, Francis DP (2022) Validation of digital twins: challenges and
opportunities. In: Feng B, Pedrielli G, Peng Y, Shashaani S, Song E, Corlu C G, Lee L H, Chew
E P, Roeder T, and Lendermann P (eds) Proceedings of the 2022 winter simulation conference.
Singapore, pp 2900–2911
47. Kibira D, Weiss BA (2022) Towards a digital twin of a robot workcell to support prognostics
and health management. In: Feng B, Pedrielli G, Peng Y, Shashaani S, Song E, Corlu CG,
Lee LH, Chew EP, Roeder T, Lendermann P (eds) Proceedings of the 2022 winter simulation
conference. IEEE, pp 2968–2979
48. ASME V&V 10-2019 (2023) Standard for verification and validation in computational Solid
Mechanics. Retrieved on 11 Oct 2023, from https://webstore.ansi.org/standards/asme/asme10
2019
49. ASME V&V 20-2009 (2023) Standard for verification and validation in computational Fluid
Dynamics and Heat Transfer. Retrieved on 11 Oct 2023, from https://www.asme.org/codes-
standards/find-codes-standards/v-v-20-standard-verification-validation-computational-fluid-
dynamics-heat-transfer
50. ASME V&V 40-2018 (2023) Assessing credibility of computational modeling through
verification and validation: Application to medical devices. Retrieved on 11 Oct
2023, from https://www.asme.org/codes-standards/find-codes-standards/v-v-40-assessing-cre
dibility-computational-modeling-verification-validation-application-medical-devices
51. Shao G, Frechette S, Srinivasan V (2023) An analysis of the new ISO 23247 series of stan-
dards on digital twin framework for manufacturing. In: Proceedings of the 2023 international
manufacturing science and engineering conference
52. Hussain M, Masoudi N, Mocko G, Paredis C (2022) Approaches for simulation model reuse
in systems design—A review. SAE Int J Adv Curr Practices Mobility 4(2022-01-0355), pp
1457–1471
53. National Academies, Sciences, Engineering, Medicine (2023) Foundational research gaps and
future directions for digital twins. National Academies Press, Washington, DC. http://nap.nat
ionalacademies.org/26894
54. Grieves M (2006) Product lifecycle management: Driving the next generation of lean thinking.
McGraw Hill, New York
55. Pérez L, Rodríguez-Jiménez S, Rodríguez N, Usamentiaga R, García DF (2020) Digital Twin
and virtual reality-based methodology for multi-robot manufacturing cell commissioning. Appl
Sci 10:3633
NASA’s Michoud Assembly Facility:
Developing a Factory Digital Twin

Greg Porter , John Vickers , Robert Savoie , and Marc Aubanel

Abstract This chapter presents a brief introduction and history of NASA’s Michoud
Assembly Facility (MAF) and current ongoing initiatives to reduce cost, improve
efficiency, and eliminate deficiencies/rework using Digital Twin (DT) technologies.
NASA’s mission to develop technologies and vehicles to literally reach for the stars
is by its nature very costly when looked at on a per-unit basis. Commercial space
flight companies, both in competition with and in partnership with NASA, are chal-
lenging NASA to explore ways to continue its unique mission while driving down
costs, particularly as these complex programs like the Artemis rocket and its compo-
nents transition from development to production. DT technologies have proven to be
effective in many commercial applications, as well as programs for the Department
of Defense, and hold substantial potential to help NASA achieve needed production
enhancements and cost reductions. The Michoud Assembly Facility is serving as a
prototype Factory DT for NASA as described in this chapter.

Keywords NASA Michoud Assembly Facility (MAF) · Digital Transformation

(DT) in Manufacturing · Environmental Monitoring and Analysis ·
Simulation-driven Manufacturing · Augmented Reality/Virtual Reality (AR/VR)
Applications · Real-time Simulation and Design · Unity and Unreal Engine
Applications · Integration of CAD and Gaming Environments · Reality Capture
Techniques · Efficient 3D Modeling · Real-time Tracking Sensors · Digital Thread
Integration · Automated Data Provisioning · Cloud Computing for Digital Twins ·

G. Porter (B)
Sev1Tech, Woodbridge, USA
e-mail: [email protected]
J. Vickers
NASA, Huntsville, USA
e-mail: [email protected]
R. Savoie
Loyola University New Orleans, New Orleans, USA
e-mail: [email protected]
M. Aubanel
Louisiana State University, Baton Rouge, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 171
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_8
172 G. Porter et al.

Kubernetes Orchestration · Pixel Streaming for Interactive 3D Rendering ·

Predictive Maintenance in Industrial Internet of Things (IIoT) · Generative AI for
Digital Twins · Automated Point Cloud Segmentation · Ontology-based
Relationships

1 Introduction

Throughout its storied history, NASA has faced three major challenges that most
organizations do not have. These challenges are: (a) the systems they create are very
expensive; (b) they make very few of these systems; and (c) the systems they make
have not been made before [1]. The difference confronting NASA, and particularly
NASA’s human space flight programs today, is the recent development of a commer-
cial space industry and the successes achieved by some of the companies in that
industry.
NASA is no longer competing only with other nations like Russia and China.
NASA now finds itself both competing with and partnering with commercial space
flight companies, particularly in Low Earth Orbit (LEO). SpaceX, for example, now
makes regular flights to the International Space Station, which is located in LEO
approximately 240 miles (400 km) above the earth. As such, it is no longer necessary
for NASA to build a separate spacecraft capacity to service LEO and compete with
these successful commercial space companies. In this case, the three issues delineated
above are no longer applicable.
NASA is now focused on the next steps where the three challenges are still appli-
cable—returning to the moon to establish a permanent lunar presence and eventually
sending the first humans to mars. The moon is approximately 240,000 miles from the
earth, which is a totally different challenge than traversing the 240 miles to LEO and
mars is 230 million miles from earth, which is a massive step even from the moon.
These are the type of challenges that fall into NASA’s mission space. The devel-
opment of the Space Launch System (SLS), the Orion Capsule and the Exploration
Upper Stage (EUS), all components of the Artemis rocket that have already circled
the moon, fit within NASA’s three key challenges. However, as SLS, Orion, EUS,
and other elements of the Artemis program move from development to production,
and as commercial space companies continue to develop more and more capabili-
ties, often in partnership with NASA, there is growing pressure on NASA to bring
down the costs of these complex systems to maintain its outreach to the moon, mars
and beyond. The DT technologies described in this chapter are one set of effectual
technologies that hold substantial promise for helping NASA to enhance production
and reduce the costs associated with continuing its mission to reach for the stars.
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 173

2 History of NASA and Michoud

The NASA Michoud Assembly Facility (MAF) in New Orleans, Louisiana, has a
long and storied history. It was originally built in 1940 by the Higgins Industries
company as a production facility for cargo aircraft and tank engines during World
War II. NASA chose MAF in 1961 as the site for the manufacturing and assembly
of the first stages of the Saturn I and Saturn V launch vehicles (Fig. 1), which were
used to send astronauts to the moon during the Apollo program. MAF also played a
key role in the construction of the Space Shuttle, producing the external tank (Fig. 2)
from the first launch in 1981 to the mission in 2011. NASA’s space shuttle fleet flew
135 missions and was the delivery vehicle to support construction of the International
Space Station.
In recent years, MAF has been involved in the development of new space explo-
ration technologies, including the Orion spacecraft (Fig. 3) and the Space Launch
System (SLS) (Fig. 4). MAF is one of the largest manufacturing facilities in the
world, with over 800 acres of land and 40 acres of manufacturing space under one
roof. Today, MAF is a vital asset to NASA’s space exploration program and to the
state of Louisiana. It is a symbol of American ingenuity and innovation, and it will
continue to play a key role in the future of space exploration.

Fig. 1 First stages of Saturn V rockets being assembled at the Michoud factory in the 1960s
174 G. Porter et al.

Fig. 2 Final external tank at Michoud Embarks on new mission of discovery

Fig. 3 Readying Orion for Flight, The NASA team at the Michoud Assembly Facility in New
Orleans has completed the final weld on the first space-bound Orion Capsule
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 175

Fig. 4 NASA Attaches First of 4 RS-25 Engines to Artemis I Rocket Stage Engineers and tech-
nicians at NASA’s Michoud Assembly Facility in New Orleans have structurally mated the first of
four RS-25 engines to the core stage for NASA’s Space Launch System rocket that will help power
the first Artemis mission to the moon

New programs have allowed MAF to evolve and adopt new physical infrastructure
and technologies where needed and in an incremental fashion to meet program needs
such as the largest friction stir welding tool in the world, built for SLS and located
in the MAF Vertical Assembly Center (Fig. 5). However, legacy facilities like MAF
pose significant challenges to the adoption of widespread new technologies, hindering
innovation and progress. Government and industry alike have failed to focus on the
challenge of how to bring innovation to legacy sectors [2].
These legacy sectors make up the majority of the US economy. There are nearly
300,000 factories in the USA, of which 90% have fewer than 100 employees.
Research indicates that it’s likely less than 10% of small and medium-sized manu-
facturing enterprises in the USA are extensively adopting emerging technologies
[3].
Legacy facilities are often designed to accommodate outdated technologies and
processes, making it difficult to integrate new equipment or infrastructure. Physical
limitations such as limited space, outdated electrical and computer systems, or inad-
equate support infrastructure, can render the implementation of new technologies
impractical or extremely difficult.
Cultural inertia is often one of the biggest challenges. Established practices and
routines within legacy facilities can create a culture resistant to change. Employees
accustomed to traditional methods may be hesitant to embrace new technologies,
even if they offer potential benefits. To overcome these challenges, organizations
176 G. Porter et al.

Fig. 5 170-foot-tall, 78-foot-wide world-class welding tool that will be used to build the core stage
of America’s next great rocket, the Space Launch System
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 177

must adopt a strategic approach to technology adoption and foster a culture of inno-
vation. To remain competitive and not become obsolete, organizations must invest in
new technology and processes and allocate sufficient resources to acquire, implement,
and integrate new technologies. Organizations can encourage employees to embrace
new ideas, provide opportunities for experimentation, reward successful innovation
efforts, and create a supportive environment. Finally, they might partner with consul-
tants, or research institutions to gain expertise, access cutting-edge solutions, and
navigate implementation challenges.
Today, these legacy challenges cause space systems development to take too long
and cost too much, and this high cost of operating in the space industry limits new
entrants and competitiveness. NASA and industry must deliver ever-more capable
products at reasonable cost. Manufacturing technology advancements must play a
critical role to improve US industry leadership in today’s hypercompetitive global
environment. The future of space manufacturing will be shaped by the digital trans-
formation and particularly by a new concept of radical innovation in design and
manufacturing called the DT.
While the terminology has changed over time, the basic concept of the DT model
has remained fairly stable from its inception in 2002 [1]. The concept of the DT
dates back to a presentation at an SME conference in 2002 by Dr. Michael Grieves
[4]. The term DT was later coined by John Vickers around the year 2010 [5], and
because of their working relationship was attached to Dr. Grieves concept.
The advancement of the DT approach has been tremendous. Today a Google
Scholar search yields more than 80,000 results and a Google search produces more
than 25,000,000 results. The DT approach has also evolved from its origins in manu-
facturing to vast applications in defense, business, transportation, science, medicine,
and a DT of the earth’s climate which is maybe the most complex system in existence.
More than anyone else, Dr. Grieves continues to shape the landscape of DTs.
To propel MAF into the future NASA has awarded Louisiana State University
(LSU) a grant to produce the first real-time virtual representation, a DT of MAF. The
DT of the facility will mirror its real-life counterpart and replace the physical trial-
and-error approaches with computational data-driven modeling to improve design,
reduce cost, and improve quality.
The MAF DT will also replace the outdated model room shown in Fig. 6.
Scale modeling or Replica Twins [6] has existed for millennia. Leonardo Da Vinci
was famous for creating intricate scale models of catapults, paddleboats, and even
mechanized robots during his life to present to the local rulers [7].
Today, the scale model for utilitarian purposes is likely obsolete. LSU will utilize
the same technology used by the entertainment industry. The LSU College of Art &
Design and the LSU Digital Media Arts & Engineering program will lead the
construction of the DT that will eventually be housed at the LSU Campus. This
involvement of the Art department comes from a lesson from the Disney Imagineers
that don’t allow the engineers to participate in the creation phase of a major attrac-
tion until the artists have produced the concept. The reason is that engineers would
compulsorily favor physics over imagination.
178 G. Porter et al.

Fig. 6 Model room at the Michoud Assembly Facility

The MAF DT will be an incredibly valuable tool, both for space industry profes-
sionals, who will be able to test a design at our facility from anywhere in the world,
as well as for educators, who will be able to bring the excitement of spacecraft
engineering to their students.

3 Benefits of the Digital Twin

The benefits that may be realized from DT, digital thread, and Augmented Reality/
Virtual Reality (AR/VR) technology depend on the use cases specific to particular
factory applications, equipment, manufacturing processes/environment, challenges,
and desired outcomes. For MAF, and particularly the production of the SLS and EUS,
the initial challenges were due to the sheer size of the MAF manufacturing facility,
primarily building 103, and the massive size of the SLS and EUS “finished products.”
The initial steps were to combine laser scanning of the facility, equipment, and tooling
with the design of the SLS and EUS to ensure the “product” could be manufactured
at MAF without impediments such as transferring parts or a completed tank through
the various openings/doorways in the factory without hitting any obstacles that could
halt production. This step has been underway for years and provided an excellent
starting point for the work described later. While this work did not constitute a true
DT, it certainly made clear the benefits and cost savings that could be achieved by
generating a DT of MAF and the technologies associated therewith. The items listed
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 179

below provide a high-level description of the potential benefits that can be realized
at MAF.
. Environmental Interrogation—floor levelness, floor load, environmental sensors,
floor shifting with the tide. The floor of MAF is concrete and covers 40 acres.
Since it was built basically at sea level near the coastal marshes of Louisiana, the
floor moves up and down slightly with the tide. Further, the floor is not uniform or
completely level. Understanding the movement of the floor and scanning the floor
to identify low spots where moisture or water might gather is invaluable when
planning the layout of the production process and understanding the environmental
effects on equipment and components being manufactured.
. Simulation—cranes, manufacturing processes—cranes are but one type of basic
tool that is needed in the manufacturing process. The location, lift capacity, main-
tenance status, etc., are all critical to the process. Cranes must also be considered
when planning the movement of large parts since they tend to be lower than the
ceiling height and could pose an impediment.
. Factory optimization—tracking parts movement over time (discussed in detail
later)
. AR for remote locations such as clean rooms—having a technician or other
personnel use AR when performing work in a hard-to-reach location, such as
a clean room, allows engineers and quality personnel to see what the technician
sees without taking the time to enter the clean room or reach the remote location.
. AR/VR-based training and work instructions—Reducing non-conformances and
rework by as much as 50%, as has been proven in other aerospace-related
production facilities [8].
. Predictive/Preventative maintenance—For cranes, tooling, welders, etc.
. NASA Metaverse—Connecting all NASA sites, monitoring in real-time rockets
or rocket sections being shipped from MAF to Kennedy or between any other
NASA Centers. Supply Chain monitoring across all NASA sites.

4 Gamification of DTs

One of the largest shifts in how we interact with digital artifacts comes down to real-
time interaction with real-world physics and data in fully realistic 3D environments.
Automobile design, simulating crash tests, architecture, medical visualization, manu-
facturing processes, and training are all being revolutionized by real-time engines.
This allows engineers and designers to manipulate and design via computers in real
time.
Two of the popular video gaming development engines are Unreal and Unity.
However, both game engines are used extensively outside of the video game devel-
opment field where these engines were initially conceived. In 2022, the total revenue
of the video game industry in the USA reached an impressive $56.6 Billion [9]. As
per a report from the Entertainment Software Association unveiled on July 10th,
2023, 65% of the American population engages in playing video games, with a
180 G. Porter et al.

resounding 96% acknowledging the perceived benefits of these interactive experi-

ences [10]. Beyond the realm of entertainment, video games have become influ-
ential in shaping designs for applications across various sectors. Platforms such as
LinkedIn [11] and renowned news publications like The New York Times [12] bear
the imprint of game concepts, fundamentally influencing human–computer interac-
tion. This phenomenon contributes to narrowing the gap between the average LSU
student and a NASA employee, both possessing a comparable level of computer
literacy and sharing a common experiential understanding of how we interact with
intricate systems.
Within the realm of DT development on gaming platforms, the capability to
engage with the physical environment in real time allows us to iterate and concep-
tualize solutions virtually, offering a means to explore scenarios before any real-
world implications. This shift in interaction has democratized tools that were once
exclusive to skilled programmers, extending their utility to professionals such as
NASA engineers without a background in software development. This allows us
to leverage familiar user experiences, particularly those drawn from commonplace
interactions with computers and video games, to lower the barrier of entry, and drive
more engagement from users.
Numerous techniques become accessible in the virtual realm that are unattain-
able in the constraints of the physical world. The virtually limitless budget afforded
by the negligible costs associated with creating a virtual rocket, in stark contrast
to the expenses incurred in manufacturing the actual rocket component, empowers
us to freely explore a multitude of ideas on a computer. This financial advantage
enables iterative processes to unfold many times for a single design, enabling collab-
orative iterations with engineers and designers. Additionally, autonomous iterations
facilitated by artificial intelligence (AI) enhance the speed of the iterative process,
surpassing the capabilities achievable by humans alone. Consequently, with AI’s
help, we find solutions to complex problems on a computer that transcends the
boundaries of human ingenuity.
A real-time 3D engine such as Unity or Unreal has another enormous advantage. It
is multi-platform and multi-user. This allows us to support a large variety of devices
interoperating together in the same simulation. For example, using the DT platform,
an operator with an AR Headset like the Microsoft HoloLens 2 could interact with
an engineer at another facility on their computer, tablet, or phone. This allows for an
operator on the floor with the 3D glasses in the facility to interact with an engineer at
another NASA location who not only sees what the operator can see but has access to
the full DT as well. This technology allows them to communicate in a rich manner that
could not be achieved by phone or messaging alone. It contextualizes the problems
they are facing and again reduces the communication gap between personnel.
A major challenge we have faced is the enormity of MAF. Both Unity and Unreal
support were skeptical about recreating a ~ 2 million square feet facility that would
contain thousands of lights, as lighting is heavy to compute. Additionally, we needed
to achieve a very high level of accuracy and details on the building and all tooling and
equipment within it. To achieve this, we use 3D laser scanning (discussed in more
detail in Sect. 5), which requires the processing of trillions of polygons in order to
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 181

represent the entire facility. These challenges led us to select Unreal 5 game engine
as the backend platform for the current iteration of the DT platform. There are a
few features in Unreal 5 that we found beneficial in overcoming these challenges:
Nanites and Datasmith imports.
Nanites replaces regular geometry within the gaming engine. Rendering using
Nanites is a fundamental shift in how to make a level, or in our case a DT, performant.
Traditionally, for 3D simulations we are “draw call bound” which results in a finite
number of polygons that the GPU can keep in memory. The ability or inability to
render a level in a game engine is the direct result of how many polygons are present.
In large environments such as the recreation of MAF, software developers would have
to use many tricks and complex techniques to lower the polygon count. However,
since Nanites is screen resolution bound and not “draw call bound,” we can have a
much larger polygonal count.
For example, if we are 1000 feet away from a piece of equipment within the DT,
the model may only consist of 10,000 polygons. But, as we approach the equipment,
the polygons increase significantly to provide an extremely realistic representation of
the equipment. A large piece of equipment can easily contain millions of polygons.
Ultimately, this allows us to achieve high accuracy and model at the highest level
of detail (LOD) possible. The engine will scale the model automatically to what is
needed to represent it realistically on screen. This is a game changer for us as it
allows us to generate a single model. With prior engines, we had to represent objects
with multiple levels of detail, as the engine now takes care of the LOD.
Based on previous experiences, we knew one of the large challenges we would
face on the project was working with engineers and architects who use CAD data
to be able to view their work in a polygonal real-time environment. However, the
compatibility between CAD and polygonal models is poorly supported. Additionally,
the techniques to design a piece of manufacturing equipment and a game differ
greatly. When designing manufacturing equipment, it is normal to create the entire
model in great detail, as a single model. In game development, modeling is typically
broken into small pieces and then placed together, analogous to building using Lego
bricks to build a model versus having a single piece be the whole model.
The solution to this problem is Datasmith. It allows us to take models created
in CAD and split them into smaller polygonal objects that we can then directly
import into Unreal 5. This closes the gap between the DT platform’s development
workflow and the CAD modeler’s workflow by increasing collaboration and reducing
the amount of rework needed. Furthermore, by incorporating these tools and data
objects into the digital thread (discussed in Sect. 6), we can fully understand and
track the relational dependencies.
By using Unreal Engine 5, we also help solve another critical problem: cyber
security. By making the DT platform widely available to many different devices such
as computers, cell phones, tablets, and AR/VR gear, and people, there is an increased
risk of a data breach. To alleviate these cyber security concerns, we are using pixel
streaming. Pixel streaming allows us to run the DT platform from remote servers
which prevents users from having to download the software. Instead, operators just
182 G. Porter et al.

connect to the DT platform with a web browser (this topic is discussed in greater
detail in Sect. 7).
Another key factor for the project is to create a user interface that is easy to use,
while allowing immersive interactions. To do this, the team is relying on sandbox-
styled games like SimCity (Maxis). The goal is to create an extremely detailed
environment with many base features to drive engagement by the operators.
Our reliance predominantly rests on sandbox-style games such as SimCity (Maxis)
for our DT development. Our emphasis lies in fostering experimentation and facili-
tating changes within the operational facility. Despite representing intricate details,
our goal is to encourage operators to experiment and craft scenarios that go beyond
the scope envisioned by the DT’s software designers. This approach not only ensures
simplicity in controlling a complex system but also poses a primary challenge for
our team: crafting a user interface that strikes a balance between ease of use and
enabling robust interaction.
Through the creation of tools designed to address envisioned use cases, we antic-
ipate diverse applications beyond the team’s initial considerations. Drawing inspi-
ration from the transformation witnessed in the video game industry, where devel-
opers released products before completion, allowing players to contribute to the final
version of the game, our approach is similarly dynamic. The DT serves as a catalyst for
this project team to significantly alter NASA’s operational landscape, contributing
to the overarching objectives of achieving lower costs, enhanced reliability, and
ultimately, the development of an improved rocket for future mars missions.

5 Unlock Business Value Through Comprehensive Reality

Capture in DTs

As mentioned in the history of NASA and Michoud section, Michoud is an older

facility and has gone through lots of changes over the years to accommodate different
space programs. As such, it was important for the project to begin with capturing the
“as-built” state to form the foundation of the DT. To do so, we have employed several
techniques which, when merged, we call reality capture. These techniques help us
deliver the NASA DT as an “as-built” replica of the physical world in a digital context,
ultimately delivering substantial business value and return on investment. The fusion
of 3D laser scanning, photogrammetry, and advanced sensing technologies elevates
the precision and utility of DTs, offering a wide array of applications across industries.
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 183

5.1 Precision Through 3D Laser Scanning

and Photogrammetry

In the pursuit of precision, imperative when building rockets, our 3D laser scanning
techniques to capture the “as-built” state of the factory, equipment, tooling, and
flight hardware are highly refined and continually tested. The utilization of laser
scanning provides a meticulous and highly accurate method for spatial data capture,
ensuring modelers have a robust foundation to work with. The incorporation of
target-based registration, which is the practice of using artificial targets, or points
and planes extracted from scan points, to align multiple scans, further refines this
process, increasing the alignment and accuracy of the captured data. This precise
dataset serves as a crucial starting point for our 3D modelers, providing them with a
detailed and reliable representation of the physical environment through the resulting
point clouds.
In looking for ways to increase efficiency, our DT platform introduces a unique
capability: the ability to display point clouds without the necessity of 3D models,
also referred to as meshes. This feature proves invaluable in scenarios where time
and cost considerations dictate the omission of a full 3D model. Point clouds within
the DT retain all the functionality and richness of a traditional 3D model, providing
a comprehensive and immersive experience. However, it is crucial to acknowledge
a tradeoff: while point clouds offer a time-saving alternative for visualizing the “as-
built” world, they require more computational power than their 3D model (mesh)
counterparts. The increased computational requirements underscore the importance
of balancing modeling intricacy with available compute resources.
Advancement in technology has paved the way for increased utilization and explo-
ration of photogrammetry, defined as the process of deriving metric information
about an object through measurements made on photographs of the object [13].
With high-definition photographs serving as the data source, it is applicable, at low
costs, to capture 3D morphology, including photo-realistic textures in high reso-
lution with large overlaps. Our modelers use this during modeling to implement
realistic textures and color palettes to create Level of Detail (LOD) 400–500 meshes
and textures. Ultimately, the goal of photogrammetry is automatic generation of 3D
models from photographs, resulting in a significant leap in reality capture capabili-
ties. However, current practices often require manual interactions such as geometric
modeling, object identification, and monitoring [14]. Despite the rapid strides in this
exciting subset of reality capture, which may one day replace most 3D modeling for
as-built objects and environments, we believe that, for the precision required in most
areas of our project, photogrammetry has not yet reached its full potential. We will
keep experimenting with and exploring the technology.
184 G. Porter et al.

5.2 Efficiency in 3D Modeling

The transition from point clouds to 3D models is where our efficiency begins to take
center stage. Using Machine Learning (ML) models, discussed in detail within the
“Artificial Intelligence, Machine Learning, and DTs” section, we can automatically
extract objects from point clouds. This process greatly reduces the time 3D modelers
must spend trying to manually separate objects from much larger point clouds. The
smaller object-oriented point cloud segmentations allow the 3D modelers to produce
highly accurate 3D models that accurately replicate the physical environment. The
conversion process of smaller, object-oriented point cloud segmentations into 3D
models is not only swift, but remarkably precise, enabling the creation of virtual
replicas that mirror reality in astonishing detail.
These 3D models within the DT platform serve as a virtual representation of the
physical environment, offering a wealth of insights for decision-making processes
and coordination; NASA and its contractors can leverage to conduct spatial analysis,
simulate various scenarios, and optimize designs. Additionally, the intricate details
captured through laser scanning later translated into 3D models empower decision-
makers with a holistic understanding of the environment, allowing for informed data-
driven choices in various aspects of project planning, execution, and maintenance.

5.3 Dynamic Spatial Awareness with Real-Time Tracking

Sensors

The integration of real-time location tracking sensors using ultra-wideband tech-

nology not only introduces a new dynamic dimension to the DT, but also paves
the way for optimizing operations within Michoud. This reality capture approach
goes beyond asset monitoring. It serves as a catalyst for comprehensive factory flow,
adherence to safety protocols, and optimal resource management.
In the context of factory optimization, the continuous tracking of assets in real time
becomes instrumental. Decision-makers, using the DT, can understand workflows
and asset movements, gaining more transparency into the intricacies of operational
processes. This historical tracking data provides a valuable foundation for optimizing
factory layouts, identifying bottlenecks, and increasing overall operational efficiency.
The DT, infused with this real-time location information, becomes a dynamic tool
for not just monitoring, but actively shaping and refining the factory environment.
The real-time location data plays another important role in ensuring swift
and accurate asset location within the expansive Michoud facility. The ability to
quickly locate objects, equipment, and tooling enhances response times, minimizes
downtime, and contributes to a more agile and responsive operational ecosystem.
Furthermore, the precision of location tracking using ultra-wideband enables a
deeper understanding of how objects move within the real-world environment. This
knowledge allows us to understand reality to accurately simulate movements within
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 185

the DT. By precisely tracking the location and movement of assets, the DT can
simulate and predict scenarios with a high degree of accuracy, offering a powerful
resource for scenario planning, predictive maintenance, complex crane lifts, and
operational optimization.
The relational linking of sensor tags to other data systems through the digital
thread is a critical aspect of this integration. It creates a flow of information allowing
the DT to tap into disparate systems containing details about the assets being tracked,
prior/future work orders, active capacities, and much more.

5.4 Environmental Insights for Sustainability and Efficiency

The integration of environmental sensors adds a layer of intelligence to DTs by

continuously monitoring factors such as temperature, humidity, and volatile organic
compounds. This real-time data is invaluable for optimizing environmental condi-
tions, ensuring occupant comfort, and identifying opportunities for sustainability
initiatives. The return on investment is manifested in energy savings, improved
workplace conditions, and a positive impact on the organization’s environmental
footprint.
The integration of environmental sensors offers another layer of reality capture to
increase the capabilities and realism of DTs. Within Michoud, we have implemented
temperature, humidity, and volatile organic compound sensors. This real-time data
further enhances the optimization of environmental conditions, uncovering avenues
for sustainability initiatives. The return on investment materializes in tangible benefits
such as energy savings, better workplace conditions, reduction in the organization’s
environmental footprint, and increased quality assurance in manufacturing processes.
In aerospace manufacturing, temperature and humidity play a significant role
in the outcomes of diverse processes like welding and chemical applications. By
tracking this data and utilizing it within ML algorithms, the DT platform becomes a
catalyst for predicting manufacturing outcomes and prescribing ideal conditions for
optimal results.
The integration of external weather data with internal sensor information further
amplifies the power of the DT. The combination allows for a comprehensive under-
standing of how external conditions impact the internal environment. By leveraging
external weather data, organizations gain insights into the broader environmental
context, enabling proactive adjustments to internal conditions based on anticipated
external changes. This interconnected approach fortifies the DT with a holistic view,
ensuring that the manufacturing environment remains adaptive and responsive to the
dynamic mix of internal and external factors.
As NASA and the rest of the world place greater emphasis on sustainability
initiatives, the environmental data generated by these sensors can help identify
areas for energy savings, enabling organizations to optimize resource utilization,
reduce energy consumption, and minimize environmental impact. All of this can be
displayed and interacted with through the DT.
186 G. Porter et al.

In essence, the efficiency in 3D modeling achieved through laser scanning tech-

niques and the use of point clouds within the DT, not only streamlines the modeling
process, but when combined with endless data from the digital thread, unlocks invalu-
able insights. Reality capture positions NASA and its contractors at the forefront of
precision-driven decision-making during the manufacturing and assembly processes,
to advance innovation, precision, sustainability, and efficiency in the dynamic realm
of space exploration and research.

6 Digital Threads: Catalysts for Transformative Data

Management

MIT defines a digital thread as a data-driven architecture that links together infor-
mation generated from all stages of the product lifecycle and is envisioned to be the
primary or authoritative data and communication platform for a company’s products
at any instance of time [15].
At the practical level, implementing the digital thread augments existing enterprise
integrations. This evolution is based on advancements within information technology,
driven by well-known and used innovations such as cost-effective and nearly limitless
computation and storage capabilities through cloud computing. Additionally, lighter,
and more reliable data transmissions, inspired by optimizations realized in internet
streaming, contribute to the seamless flow of information between systems. This
contemporary landscape is further shaped by faster and stronger databases, new ways
of controlling versions from disparate systems, affordable and dependable sensors,
and the application of sophisticated techniques in physical object scanning and image
processing.
The integration of these innovations not only strengthens the existing data infras-
tructure but also enriches the product development process measurably. The digital
thread allows for comparison, including visualizations, of product data across diverse
lifecycle stages. For example, it facilitates the comparative analysis of “as-built”
digitalized tolerance measurements against the original design in a computer-aided
design (CAD) file. This integration introduces closed feedback loops, contributing to
a more comprehensive understanding and optimization of the product development
lifecycle.
Value also arises from the automated provisioning of updated authoritative data
sources, making the digital thread itself the authoritative source. It allows data to
stay within its source system, exposing only necessary metadata until the actual
data is needed. This empowers subject matter experts to retain control over their
information. Unlike traditional data workflows requiring resource-intensive Extract,
Transform, Load (ETL) workflows to replicate data into a centralized data warehouse,
the digital thread renders these processes obsolete by allowing data to reside within
its originating system. This streamlined approach conserves organizational resources
and enhances agility and responsiveness.
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 187

Beyond data control, digital threads have established themselves as the corner-
stone in DT architecture by facilitating the near real-time synchronization of digital
system models and their corresponding DTs. Establishing connections between
disparate systems fosters a relational understanding of the data, systematically
breaking down data silos and providing organizations with a comprehensive view
that leads to novel and insightful observations. The relationships between systems
can be created through ontologies, which represent a shared, explicit specification of
a conceptualization of the domain of knowledge [16]. It provides a formal specifi-
cation of the vocabulary of concepts and the relationships among them, in a domain
of interest [17].
In the diverse ecosystem of NASA, comprising vendors, contractors, and tenants,
each entity operates with distinct software, processes, and scopes of work. To deploy
an effective digital thread in such an ecosystem, it must be platform agnostic, allowing
entities to continue their operations as usual. Complex systems can be interconnected
through innovative tools such as the digital thread, utilizing modern-day Application
Programming Interfaces (APIs). These APIs enable seamless interactions with the
digital thread, facilitating native integration and data exchange.

7 DT Accessibility

Traditionally, operating an application with 3D graphics or models required the

operator to run a system with a high-end Graphics Processing Unit (GPU). This has
limited the number of users able to use and interact with the application. On this
project, we sought to solve this problem, enabling us to deliver a massively complex
3D render involving point clouds, models, and high-fidelity textures from anywhere
within a factory while sticking true to being platform agnostic. We have successfully
done so using technologies such as multi-tiered cloud computing, Kubernetes, and
pixel streaming.

7.1 Cloud Computing: Flexibility, Cost-Efficiency,

Scalability, and GPU Acceleration

Cloud computing has become a linchpin for businesses seeking flexibility, cost-
efficiency, and scalability in their IT infrastructure. One pivotal advancement in cloud
computing is the availability of GPU resources, significantly enhancing the perfor-
mance of applications that require intensive graphical processing. This is particularly
beneficial for workloads such as ML, simulations, and realistic rendering of DTs. By
leveraging cloud services with GPU support, organizations can offload the burden
of managing the procurement and deployment of specialized hardware, ensuring
that resource-intensive tasks are executed efficiently. By leveraging cloud services,
188 G. Porter et al.

our project has been able to offload the burden of managing physical hardware,
allowing us to focus on developing automated and highly scalable DT platforms,
interconnected digital threads, and near real-time 3D model data catalogs.
The pay-as-you-go model of cloud computing ensures cost-effectiveness, since
we only have to pay for the resources we consume. Moreover, the scalability of
cloud resources enables us to effortlessly handle fluctuations in demand by using
autoscaling and rightsizing best practices for modern-day architectures, ensuring
optimal performance during peak times, and cost savings during periods of reduced
activity. Cloud computing services have enabled us to innovate rapidly and remain
agile in today’s dynamic digital landscape. Furthermore, by using security groups,
we implicitly deny all access and explicitly allow access, giving us granular control
over infrastructure and applications.

7.2 Kubernetes: Streamlining Container Orchestration

for Scalability and Reliability

Kubernetes, an open-source container orchestration platform, has revolutionized the

way applications are deployed, scaled, and managed. One of its primary bene-
fits is its ability to automate the deployment, scaling, and operation of appli-
cation containers, providing a consistent and resilient environment while being
platform-agnostic and open-sourced. Kubernetes facilitates the efficient orchestra-
tion of containers, ensuring optimal resource utilization and seamless scaling to meet
varying workloads. With features like automated load balancing and self-healing
capabilities, Kubernetes enhances the reliability and availability of applications,
making it a cornerstone technology for modern, cloud-native, hybrid, or on-premises
architectures.

7.3 Pixel Streaming: Interactive 3D DT Anywhere, Anytime

Unreal Engine Pixel Streaming introduces a paradigm shift in content delivery by

enabling the streaming of interactive 3D content directly to a user’s device. This
technology allows resource-intensive 3D graphics to be rendered on a server running
in the cloud or on-premises and streamed to end-users in real time. This provides many
benefits such as eliminating the need for users to possess high-end hardware to run
demanding applications, facilitating collaboration on complex 3D environments in
real time, and presenting 3D applications in web-based applications. Pixel streaming
has empowered us to deliver an immersive DT to a broader audience, irrespective of
device specifications, fostering a more inclusive and accessible DT platform.
Unreal Engine’s Pixel Streaming capabilities represent a revolutionary shift in
content delivery, enabling the real-time streaming of interactive 3D content directly
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 189

to users’ devices. This technology allows resource-intensive 3D graphic applications

to be processed on a server and dynamically streamed to end-users. Notably, Pixel
Streaming provides many advantages, including but not limited to, eliminating the
need for users to possess high-end hardware to run demanding 3D applications and
facilitating real-time collaboration on complex 3D environments in real time.
Adding to its versatility, Pixel Streaming integrates with modern website design
frameworks like React and Next.js. This integration provides an added layer of flex-
ibility, allowing us to embed our interactive 3D DT into a modern web application.
This not only enhances the user experience but also opens new possibilities for
creating new engaging tie-ins with the DT platform like dynamically streaming data
from the digital thread. The intersection of Pixel Streaming and web design method-
ologies represents a convergence of worlds, fostering a balanced blend of immersive
3D content and the latest trends in web development.
Beyond these capabilities, Pixel Streaming acts as a gateway to collaborative
experiences within the DT through simulations and augmented reality (AR) based
scenarios. It facilitates multiple users interacting with each other, whether connecting
through AR devices, phones, or computers. This breakthrough allows users to share
and jointly manipulate 3D content in real-time encouraging collaborative decision-
making and enhancing the overall user engagement. Seamlessly integrating all these
capabilities mixed in with federated identities, coupled with the digital thread as the
single source of truth, allows us to ensure the delivery of an immersive environment,
utilizing a security best practice known as least privilege.

7.4 Putting It All Together

The convergence of Kubernetes, cloud computing, and Pixel Streaming lays a robust
foundation for DTs. In this advancing technological landscape, the integration of
these tools is proving to drive innovation across not only NASA and aerospace indus-
tries alike, but also education, manufacturing, energy, defense, real estate, marketing,
and many others.
By containerizing Pixel Streaming and deploying it on Kubernetes within a cloud
environment, we provide a consistent and portable DT that is buildable, testable,
and deployable through continuous integration/continuous deployment (CI/CD)
pipelines. Deploying on Kubernetes introduces unparalleled scalability, allowing us
to efficiently handle varying workloads and fluctuations in user demand, even scaling
down to zero resources when no one is using the DT platform. This dynamic scal-
ability, coupled with the inherent accessibility of cloud resources, ensures optimal
performance regardless of geographical location or personal hardware specifications.
Ultimately, this leads to cost reduction by optimizing resource utilization and elim-
inating unnecessary expenses. The orchestrated collaboration between Kubernetes,
cloud computing, and Pixel Streaming reflects not only the current but also the future
state of technology, underscoring a vision where interactive, visually compelling
190 G. Porter et al.

content is universally accessible, affordable, and responsive to the evolving demands

of a global user base.

8 Maximizing Returns Through Artificial Intelligence,

Machine Learning, and DTs

In the dynamic landscape of DTs, the strategic integration of AI and ML not only
revolutionizes operational patterns but also serves as a jumping point for maxi-
mizing return on investment. As organizations increasingly harness the power of
DTs, the infusion of intelligent technologies amplifies their transformative potential.
This section delves into key applications within the MAF DT, focusing on our DT
assistant, point cloud segmentation, and predictive maintenance, all underpinned by
a central theme—the realization of substantial returns on investment. By exploring
how AI and ML elevate decision-making, streamline workflows, and enhance predic-
tive capabilities, we unveil the many ways in which NASA stands to gain significant
value and efficiency from their investments in these cutting-edge technologies.

8.1 Digital Twin Assistant and the Digital Thread

The DT assistant, intricately tied to the digital thread, represents a groundbreaking

development. Leveraging large commercially built Large Language Models (LLMs)
through transfer learning, the DT assistant gains a profound understanding of NASA
data beyond its initial training data, which is a large corpus of publicly accessible
internet data. The use of transfer learning, where we only retrain the last layer of the
model, democratizes access to sophisticated language models, enabling the reuse of
expansive knowledge for enhanced contextual understanding. This saves significant
time and cost when developing a custom solution involving LLMs.
Conversational scoping is a crucial aspect, ensuring the DT assistant remains
focused on the specific data we have trained it on. This precision in scoping allows
for accurate and relevant responses, aligning with data privacy and security consid-
erations. Additionally, the assistant demonstrates a unique ability to comprehend
relationships between disparate data systems through the digital thread. This inter-
connected intelligence enhances the DT’s capacity to provide nuanced information
and insights, allowing for more informed decision-making across diverse domains.
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 191

8.2 Point Cloud Segmentation

Automated point cloud segmentation has emerged as a pivotal application within our
DT project, specifically in expediting the 3D modeling workflow. By using state-
of-the-art Learnable Region Growing for Class-Agnostic Point Cloud Segmentation
ML model, we have been able to automate the segmentation of point clouds. By
transforming extremely large, dense point cloud data into structured segmented point
clouds we have significantly reduced the time and effort traditionally associated with
manual segmentation.

8.3 Predictive Maintenance

The convergence of AI, ML, and DTs facilitates predictive maintenance strategies
that revolutionize Industrial Internet of Things (IIoT) applications. Capturing IIoT
data over time allows for a deep understanding of patterns within the data, enabling
the prediction of potential issues before they occur. This capability extends beyond
routine machinery maintenance, delving into complex scenarios such as predictive
crane maintenance and assessing the effects of environmental factors like temperature
and humidity on critical processes like welding and chemical applications.
In summary, the harmonious integration of AI and ML technologies within DTs
transforms these virtual replicas into intelligent, dynamic entities. The DT Assistant,
point cloud segmentation, and predictive maintenance represent key applications,
showcasing the immense potential for enhanced decision-making, operational effi-
ciency, and innovation across diverse industries. This synergy underscores the pivotal
role of AI and ML in shaping the future trajectory of DT technologies.

9 Conclusion

In this chapter, we have covered the intricacies of the NASA Michoud Assembly
Facility (MAF) DT project, a groundbreaking initiative at the forefront of precision
engineering and exploration. As we conclude our exploration of the five key dimen-
sions: gamification of DTs, comprehensive eality capture of DTs, digital threads,
DT accessibility, AI, ML, and DTs of this comprehensive endeavor, it is evident
that the integration of cutting-edge technologies is reshaping the landscape of space
exploration and manufacturing.
The historical context provides a foundation for understanding the evolution of the
Michoud facility and NASA’s rich legacy. This historical perspective underscores the
significance of adapting traditional facilities to the demands of contemporary space
programs, setting the stage for the transformative journey into the digital realm.
192 G. Porter et al.

Reality capture techniques, comprising 3D laser scanning, photogrammetry, and

real-time tracking sensors, lay the groundwork for an unparalleled level of precision
in creating the DT. The commitment to capturing the “as-built” state of the physical
world reflects not only an attention to detail but a strategic investment in technologies
that deliver substantial business value and return on investment.
Efficiency in 3D modeling, dynamic spatial awareness, and environmental insights
contribute to the active and adaptive nature of the DT. Decision-makers are empow-
ered with tools for spatial analysis, scenario simulation, and design optimization,
fostering a holistic understanding of the environment. The introduction of real-time
location tracking sensors, coupled with environmental sensors, amplifies operational
efficiency, safety, and sustainability within the Michoud facility.
The concept of the digital thread emerges as a unifying force, seamlessly inte-
grating information across the product lifecycle. Cloud computing, Kubernetes, and
pixel streaming introduce a new era of accessibility, flexibility, and cost-efficiency.
The democratization of 3D graphics ensures that the DT is not just a high-end
GPU endeavor but a universally accessible and inclusive platform, aligned with the
principles of modern technological landscapes.
In the final dimensions explored, we witness the strategic integration of AI and ML
technologies, positioning the DT as a dynamic entity capable of substantial returns on
investment. The DT Assistant, point cloud segmentation, and predictive maintenance
showcase the transformative potential of these intelligent technologies, enhancing
decision-making, operational efficiency, and innovation across diverse industries.
As we conclude this chapter, it is clear that the NASA Michoud Assembly Facility
DT project is not merely a technological endeavor but a pioneering journey into the
future of precision engineering and exploration. The intersection of historical legacy,
eality capture, digital threads, and intelligent technologies propels NASA into a realm
where informed decisions, efficient workflows, and sustainable practices converge.

References

1. Grieves M, Vickers J (2017) DT: mitigating unpredictable, undesirable emergent behavior in

complex systems. In: Kahlen F-J, Flumerfelt S, Alves A (eds) Trans-disciplinary perspectives
on system complexity. Springer, Switzerland, pp 85–114
2. Bonvillian W, Singer PL (2017) Advanced manufacturing: the new American innovation
policies. The MIT Press, Cambridge, Massachusetts, ix, 401 pages
3. Vickers J (2022) Advanced manufacturing transformation. In: Aerospace & Defense Manu-
facturing Industry Report 2022. SME, p 3
4. Grieves M (2002) Completing the cycle: using PLM information in the sales and service
functions [Slides]. In: SME management forum, Troy, MI
5. Piascik R et al (2010) Technology area 12: materials, structures, mechanical systems, and
manufacturing road map. NASA Office of Chief Technologist
6. Grieves M, Hua E (2024) Defining, exploring, and simulating the DT metaverses. In: Grieves
M, Hua E (eds) DTs, simulation, and metaverse: driving efficiency and effectiveness in the
physical world through simulation in the virtual worlds, Forthcoming. Springer, Heidelberg
7. Anonymous (2018) The history of plastic scale modeling [cited 2023 December 1, 2023].
Available from: https://www.megahobby.com/blog/the-history-of-plastic-scale-modeling/
NASA’s Michoud Assembly Facility: Developing a Factory Digital Twin 193

8. Fink C (2019) Enterprise AR use cases. Available from: https://www.forbes.com/sites/charli

efink/2019/02/26/enterprise-ar-use-cases/amp/
9. Clement J (2023) Annual Revenue of the U.S. Video Game Industry 2016–2022. Avail-
able from: https://www.statista.com/statistics/249996/annual-revenue-of-the-us-video-game-
industry-by-segment/
10. Anonymous (2019) Essential facts about the computer and video game industry. Available from:
https://www.theesa.com/wp-content/uploads/2019/05/ESA_Essential_facts_2019_final.pdf
11. Lindemann T (2019) How LinkedIn uses gamification to boost engagement. Avail-
able from: http://thomas-lindemann.com/gamification-en/how-linkedin-uses-gamification-to-
boost-engagement/
12. Tribbey C (2018) Viacom, NYT share how they gamify the data experience. Avail-
able from: https://www.mesaonline.org/2018/07/31/viacom-nyt-share-how-they-gamify-the-
data-experience/
13. Mikhail EM, Bethel JS, McGlone JC (2001) Introduction to modern photogrammetry. John
Wiley & Sons
14. Qin R, Gruen A (2021) The role of machine intelligence in photogrammetric 3D modelling—an
overview and perspectives. Int J Digital Earth 14(1):15–31
15. Singh V, Wilcox KE (2018) Engineering design with digital thread. AIAA J 56(11):4515–4528
16. Gruber T (1993) A translation approach to portable ontology specifications. Knowl Acquis
5(2):199–220
17. Ganon M (2007) Ontology-based integration of data sources. In: 2007 10th international confer-
ence on information fusion, Quebec, QC, Canada, pp 1–8. https://doi.org/10.1109/ICIF.2007.
4408086
Distributed AI Modeling and Simulation
for Smart Airport Digital Twin
Applications

Kostas Alexandridis , Soheil Sabri , Jeff Smith, Bob Logan,

Katalin Bartfai-Walcott, and Doug Migliori

Abstract This chapter overviews modeling and simulation methodology for smart
airport Digital Twin metaverse applications, emphasizing the development of a
distributed agent-based modeling implementation framework. It outlines current
approaches and applications of modeling and simulation concerning optimization,
sustainability, and digital transformation, along with theoretical, statistical, and
empirical approximations to model validation and accuracy assessment. It also pro-
vides a rudimentary Agent-Based Modeling (ABM) categorization and classifica-
tion of current models. We provide a key perspective to airport planning model
development, including resource allocation, workflow design principles, and process
optimization. An example reference and design of distributed and intelligent agent
architecture for a smart airport is provided, along with the design elements for agent
allocation, communication, and orchestration mechanisms for system architecture.
Finally, we provide a series of inferences, insights, and principles to guide the design,

K. Alexandridis (B)
Orange County Public Works, OC Survey Geospatial Services, Santa Ana, CA, USA
e-mail: [email protected]
S. Sabri
Urban Digital Twin Lab, School of Modeling, Simulation and Training,
University of Central Florida, Orlando, Florida, USA
e-mail: [email protected]
J. Smith
Sierra Nevada Corporation, Nashua, NH, USA
e-mail: [email protected]
B. Logan
Rockport Software, Slough, England, UK
e-mail: [email protected]
K. Bartfai-Walcott
Ambient Enterprises Inc., California, USA
e-mail: [email protected]
D. Migliori
Event Driven Systems, Irvine California, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 195
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_9
196 K. Alexandridis et al.

development, and implementation of distributed AI and agent-based modeling and

simulation methodologies for smart airport and nested smart cities’ methodological
approaches.

1 Introduction

A growing body of practical and scientific work recognizes the importance of Agent-
based Modeling (ABM) and simulation in enhancing transportation operations. For
example, passenger flow modeling is one of the widest applications of ABM in
terminal models, including railways and airports. This approach has been used for
many use cases, such as improving non-aeronautical revenue, passenger satisfaction,
hazard mitigation and evacuation, and energy efficiency [1, 2]. However, there are
limited scientific reports on system architectures that incorporate airport planning
and operations into a holistic digital environment. Furthermore, the current top-down
management approach in many airports and other transportation hubs overlooks the
human and social factors as essential dimensions in the operational planning and ser-
vice allocation [3]. Given a progressive increase in the deployment of sensors, mon-
itoring equipment, and high-performance computing technologies, the traditional
modeling, simulation, and prediction approaches could be enriched with real-time
and streamlined data.
This chapter concerns optimized ABMs calibrated with real-time data. The cur-
rent state-of-the-art ABM supports a distributed AI-enabled metaverse, facilitating
transparency and flexibility of management and operations in complex environments
[4]. This method can also enable an adaptive system of systems (SoS) where stake-
holders collaborate and understand the complex implications of their decisions and
action scenarios.
We propose a distributed AI-based modeling and simulation architecture for smart
airports. To do this, the chapter first explores the state-of-the-art applications of ABM
in airports. We will explore the current applications of modeling and simulation and
ABM in different dimensions, including human systems, physical and operational
aspects, and sustainability. We pay specific attention to the Modeling and Simulation
in transportation hubs in a subsection. Then, we investigate the validation approaches
such as spatial, temporal, systemic, and behavioral. These will be translated into our
proposed architectural principles. The next section describes an airport planning
model, the definition of its components, rules, resource criteria, and usage. Using
the content of this airport planning model, we propose a smart airport ABM design,
and then we provide the visual metaverse aspects of an intelligent Digital Twin
framework for smart Airports in the next section. Finally, the chapter provides the
concluding remarks and sets an agenda for the development and implementation of
a new generation of ABM integrated into the Digital Twin architecture to enhance
Airport Planning.
Distributed AI Modeling and Simulation for Smart Airport … 197

1.1 Current Applications

Given the complexity of airport operations, this chapter focuses on three interre-
lated dimensions: (a) the human dimension (passenger flow); (b) the physical and
operational dimensions (terminals and airlines), and; (c) the sustainability dimension
(energy efficiency). In this section, we review and synthesize the current state of the
art in modeling, simulation, and virtual representation of airport operations, as well
as social and behavioral theories and validation methods to identify the progress and
limitations in the applied and scientific environments.

1.1.1 The Human Dimension

Recently, a group of studies focused on behavioral studies and incorporated humans’

discretionary activities (e.g., shopping duty-free) into the passenger flow models
[1, 5]. From a business perspective, these models improved airport terminal analytics
to support understanding the non-aeronautical revenue within the aviation industry.
For instance, these models can improve the airports’ operational strategies that reduce
passenger queue time or increase passenger free time, potentially enhancing airports’
terminal performance through improved efficiency, increased revenue, and reduced
cost [1]. Other benefits of such models are investigating the movement of a large
number of people within confined areas and linking to the definition of impacts that
can be seen in terms of time spent in each processing unit, preference of discretionary
activity, and level of service (LOS) in the processing areas [6].
Furthermore, the discretionary attributes combined with social theories, such as
social force, improve our understanding of passenger behavior during an emergency
or hazard mitigation [7]. Specifically, after the calibration of these ABM models,
the outputs can be used to improve the design elements of pedestrian facilities and
egress routes [8]. It will provide an opportunity to explore different scenarios, such
as using portable “obstacles” to stabilize flow patterns and make them more fluid in
a real-time and 3D environment.
Another important factor related to the human dimensions when considering air-
port-related systems is that human behavior in routing passengers and passenger
flows is non-deterministic, often volatile, and highly uncertain. Unlike purely phys-
ical systems or robots that can be logically routed and queued, human dispositions,
behavior, attitudes, and satisfaction play a key role in routing decisions. Uncertainty,
unpredictability, and latent patterns of collective behaviors result in challenges, prob-
lems, and decision-related systemic failures [8–10].

1.1.2 Physical and Operational Dimensions

Several studies indicated the importance of physical and operational strategies for
passenger flow control. For example, increasing the service process capacity and
198 K. Alexandridis et al.

social distancing during the COVID outbreak informed a better intervention and
decision-making process in major transportation hubs [11]. Examining the inbound
passenger flow lines through the average inbound time of passengers from each
node (parking to the entrance and thereafter) and the average queue length for each
kind of passenger service equipment facilitates gaining an insight into the terminal
layouts [12]. The latter can also be modeled as the average service/processing time
of each service, including passenger service equipment, while average queue length
and wait times can be derived from combining service times and passenger volumes
interactively [13].
Other studies looked into integrating pedestrian facilities planning and staff
assignment for transfer capacity, transfer average time, and level of service [14].
The passenger flow line regulated the entrance and transition gates, such as bidirec-
tional automatic fare gates (BAFGs), and avoided passenger congestion [15]. Given
the time-varying passenger demands, this operational problem could be improved
through real-time/near real-time data.
Undoubtedly, the airlines, as main stakeholders, play a crucial role in airport
operations, and their operational model contributes to the smoothness of passenger
flow. The service process capacity and staff assignments mentioned above could
affect airlines’ operational strategies. Airline’s approach to passenger boarding has
implications for passengers’ satisfaction, the operational timing, and the energy used
to air condition the boarding bridges [16]. As such, the physical and operational
dimensions are interrelated to human activities and sustainability.

1.1.3 Sustainability

Airport terminals consume more energy than normal public buildings on average
[17]. Studies indicated that three factors are correlated to the airport terminal’s energy
consumption: (a) passenger flows, (b) meteorological parameters, and (c) supply fan
frequency (for the zonal airport terminal HVAC system) [18]. Given these factors,
the role of ABMs in understanding the environmentally sustainable performance of
airports is significant. We argue that the energy consumption of airport terminals
poses a substantial challenge to the carbon neutrality policy due to the high energy
demand and round-the-clock operation of these terminals. Integrating geospatial
data and analytics into the ABM allows the development of the spatio-temporal
passenger distribution model to describe the transient passenger distribution pattern
of an airport terminal [19, 20]. The outputs of such models are used as input into the
energy simulation tools to estimate the energy consumption of airport terminals in
different scenarios [21].
Furthermore, passenger flow plays a vital role in the terminal’s energy consump-
tion and indoor environmental conditions [22]. A prediction model of passenger
flow, developed by [22], confirmed the operation optimization and terminal’s lay-
out design to address energy efficiency. Alternatively, energy consumption for each
service level can be aggregated based on overall service use [13].
Distributed AI Modeling and Simulation for Smart Airport … 199

1.1.4 Modeling and Simulation for Transportation Hubs

Recent research has highlighted the value of hybrid models in transportation hubs
and, specifically, airport studies [23]. To improve the reliability of models, the com-
bination of modeling and simulations and the integration of emerging technologies
such as Machine Learning (ML) and Artificial Intelligence (AI) for optimization and
predictive analysis are suggested.
For example, the passenger movements, multi-gate and multi-destination nature,
passenger service processing nodes, and their operational patterns in airports justified
the hybrid ABM and discrete event system (DES) models to develop a multi-function
simulation tool to help inform decision/policymakers in different operational phases
of planning, design, development, and implementations for addressing issues like
public health emergencies and emergency evacuations [11].
Example application integrated agent-based simulations and machine learning
methods, such as random forests, enable broader model optimization and estimation
robustness for transportation hub’s facility planning and staff assignment [24]. How-
ever, some challenges remained unresolved for the hybrid models due to the limited
availability of simulation platforms, high demand for computational resources, and
limitations in validation and optimization [11].
In macro-to-micro simulation models as a group of hybrid models, agent micro-
behaviors and spatio-temporal movements are often combined to understand and
model broader system dynamics. For example, in an airport model [15], passenger
micro-behaviors and movements aid the understanding of system-wide efficiency
patterns and parameters while enabling optimized configuration or efficient arrange-
ment of system resources.
In addition, models of flow dynamics in terms of spatial and temporal movements
and shifts are categorized as prediction models. Of particular importance is under-
standing optimization and flight arrangements to predict passenger flows in airport
terminals or using the reverse problem formulae, using predictive passenger flows
across space (terminal sections, gates, security points) and time (peak times, off-
peak times) to optimize resource and system provisions or estimate carrying/critical
capacity flows for the terminal system.

1.1.5 Validation Approaches

Often, model validation and accuracy issues fall within four related categories and
groups. Specifically, these can be (a) spatial or geostatistical, (b) temporal, (c) sys-
temic or engineering, and (d) behavioral or intelligence-related. Validation and reli-
ability represent a cornerstone for the evolution of metaverse in the future mobility
[25].
200 K. Alexandridis et al.

Spatial/Geostatistical

Normal statistical model validation and accuracy assessment concern what and to
what degree something occurs or emerges from the model behavior. Geostatistical
model validation approaches not only must incorporate standard statistical methods
but additionally must include where and to which scale such patterns and model
behaviors occur [26]. Issues of spatial accuracy often involve spatial statistics about
spatial autocorrelation, spatial heteroskedasticity, and scale variance (or invariance).
For example, validation patterns that pass accuracy testing at 1:100,000 scales might
fail at 1:10,000 scales and vice versa, rendering scaling up or down for model dynam-
ics problematic or completely fallacious. Modeling and simulation patterns of behav-
iors and results must often be able to stand scaling-up and down approaches and thus
require multi-resolution validation data or data sources that vary in spatial intensity
or at least sampled across a varying degree of spatial resolution.

Temporal

Assessing and validating temporal modeling and simulation dynamics requires a

measurement of scale and uncertainty. Temporal scales range from historical, near-
time, real-time, near-term, to long-term patterns. Each scale often requires a different
regime or battery of model validation methods and tools. More often than not, tempo-
ral model dynamics are interwoven and closely coupled with related spatial dynamics
in ways that are not easily distinguishable or cannot be readily decoupled. Spatiotem-
poral modeling and simulation models often require multi-dimensional model val-
idation and accuracy assessment, spatio-temporal stratified sampling, and datasets
(e.g., data cubes or multi-dimensional data samples). In other cases, temporal mod-
eling dynamics exhibit deep uncertainty, are riddled with incomplete information,
or lack historical data support for statistical validation. In such cases, model assess-
ment and validation methods can include subjective and latent statistical methods of
assessment or conditional/assumptional modeling assessment (e.g., finding critical
conditions where model assumptions fail or lead to unexpected results or patterns).

Systemic/Engineering

Examples of these validation and accuracy challenges include assessing system

variable inclusion, issues of equifinality [27], such as issues of resource conflicts,
and mutually exclusive and competitive system dynamics. Validation approaches
addressing these challenges must not only have a solid foundation on rich and diverse
data sources but also take into account the integrated nature of system behaviors,
account for often amphi-directional forces and drivers of change, and whole-life-
cycle engineering aspects from constructive to generative, to emergent mechanisms
and dynamics.
Distributed AI Modeling and Simulation for Smart Airport … 201

Behavioral/Intelligence

In many cases, micro-finite data patterns on individual behaviors are often desired and
needed to validate model estimations and results. For example, Scala et al. [28] used
historical data and data analytics methods to validate their modeling and simulation.
They generated a passenger “Activity-Travel-Diary” model from different sources,
such as passenger’s data (personal mobile phone, apps), airport data (airport WiFi,
GPS, scanning facilities), and flight information (flight schedules and gate alloca-
tion) and successfully validated the simulation results. Creative uses of existing data
and multi-source data pluralism can provide valuable insights into model behaviors
and predictive patterns emerging from modeling and simulation approaches. On the
other hand, behavioral and cognitive patterns and data cannot be used strictly to
establish causal predictive behaviors and require additional heuristic, probabilistic,
or propabilogic assessments and model validation methodologies. Agent-baMicro-
finite particularly sensitive in terms of statistical validation methods, as their design
and definitions encapsulate and embed a certain degree of dynamism, randomness,
and uncertainty. Heuristic approximations and assessment of realities also require
non-static methods of assessment and repeated measures such as statistical boot-
strapping, Monte Carlo simulations, or simulation ensemble methods.

1.2 Architecture Principles

1.2.1 Two Categories of ABM

This study focuses on two categories of agent modeling: (a) phenomena-based mod-
eling; and (b) exploratory modeling [29]. In phenomena-based modeling, one uses
agent rules to create known phenomena represented as a reference or aggregate pat-
tern. An example of this with respect to an airport planning model is to change a
traffic distribution pattern that may be derived from various sources including Airport
Operator or Air Traffic Management.
In exploratory modeling, one creates a set of agents, using their behavior to explore
emergent patterns. An example of this exploratory modeling with respect to an airport
planning model is given in [30] where one simulates an airport with passenger agents
passing through several airport zones, varying parameters such as boarding gates,
shops, and arrival patterns, to observe the emerging behaviors.
We expect to use both methods, with the exploratory modeling providing the basic
agent framework and parameters and the phenomena-based method used to replace
the simplified distribution patterns used in the exploratory model.
As one refines a model with either modeling category, conformance to agent
modeling standards such as the Foundation for Intelligent Physical Agents1 (FIPA),

1 www.fipa.org.
202 K. Alexandridis et al.

its Agent Communication Language2 (ACL), (fipa.org/repository/aclspecs.html) and

the burgeoning OMG Agent and Event Metamodel [31]) will connect phenomena or
exploratory-based agent modeling exploration in a more formal Model Based System
Engineering (MBSE) architecture.

1.2.2 Metaverse

Recent digital innovations and advancements, such as Digital Twins, metaverse,

and parallel systems, provide an ecosystem for mirroring real-world entities, their
dynamics, and interactions. These technologies have been proposed and customized
for adoption in several domains, such as the gaming industry, art, communication,
fashion, and education [32]. Recently, these technologies received significant atten-
tion among scholars in applied science and engineering sectors, such as manufac-
turing [33], energy [34], infrastructure planning [35, 36] and urban transportation
[37, 38].
However, most of the current studies and technical reports remained at the concep-
tual level and provide general descriptions of the values of immersive and interactive
environments for more sustainable, inclusive, and economically viable decision-
making. Technologies such as extended reality, sensor-based object recognition,
location tracking, and haptic augmented reality systems in the metaverse interactive
environment are regarded as potential solutions for complex systems like airports,
but more empirical studies are required to evaluate their feasibility [39]. Because of
their complexity, airports will require a wide variety of Digital Twins from different
sources and a wide variety of use cases. This will require interoperability of these
Digital Twins to work together seamlessly. A Digital Twin metaverse that supports
this capability will need to have certain defined features and capabilities that will
need to be articulated [40].

1.2.3 Challenges

Most current approaches, however, have failed to provide a robust, reliable, and
scalable solution to incorporate real-time data in the calibration, validation, and sen-
sitivity of tools. The existing data collection methods often lead to biased data [41].
Most data used to calibrate the models are historical or experimental. Accordingly,
agent-based models such as passenger flow models are highly simplified. The lack
of data led to the introduction of many assumptions to the model, which decreased
their reliability. In addition, in many cases, the airport terminals are considered indi-
vidual entities, whereas these buildings are connected to broader and mass build-
ings/infrastructures and catchment areas in their precinct and city (e.g., parking, city
transportation).

2 fipa.org/repository/aclspecs.html.
Distributed AI Modeling and Simulation for Smart Airport … 203

The complexity of airport systems cannot be addressed by linear and ad hoc

modeling and simulation. One of the issues for this fragmentation is the lack of data
standardization and interoperability. This might be due to the lack of an ontological
framework across different actors and stakeholders in airport operations.
As such, emerging technologies such as Metaverse and Digital Twins, which
improve the understanding of human-environment interactions, play a crucial role in
airport operations and decision-making.

1.2.4 Proposed Reference Architecture

We propose integrating complex-adaptive and multi-agent-based systems into the

Digital Twin framework to focus on human-environment interactions. We also
emphasize the significant role of geospatial intelligence, including data processing,
advanced analytics, and visualization, in improving the accuracy and granularity of
models.

1.2.5 Multi-Agent-Based Systems

Multi-agent architectures, in general, are ideal for linking Digital Twin use cases.
The multi-agent network on the “digital” side can be arranged to resemble the “real”
side structure. Each agent on the “digital” side can freely communicate with the
sensors on the “real” counterpart. Parallels between the Digital Twin and agents are
given in Table 1.
In our case, the phenomena-based modeling part approximates actual airport traf-
fic with known patterns and varying initial parameters (e.g., in and out traffic arrival
rates, number of agents, boarding gates, shops, check-in boxes, baggage belts, and
station lag times). The exploratory-based modeling part tests the consequences of
using different interaction protocols, varying the same parameters. Many commer-

Table 1 Relationships between Digital Twin and agent entities

Digital Twins (DT) Agents
.•An organized set of digital models .•Evolved and planned behavior using agents
representing a real-world entity designed to with singular functions to address specific
address specific issues/uses issues and uses

.•Updated in relation to reality, with a .•Updated in relation to reality, with a

frequency and precision adapted to issues and frequency and precision adapted to issues and
uses uses

.•Equipped with advanced operating tools, .•

Equipped with the ability to monitor,
including the ability to understand, analyze, understand, analyze, predict, and optimize
predict, and optimize
204 K. Alexandridis et al.

cial frameworks rapidly create agent applications using the Foundation for Intelli-
gent Physical Agents (FIPA) and Agent Communication Language (ACL) standards,
using tools, e.g., AgentSpeak, JACK, JADE, GAMA, Mason, Repast, and NetLogo.
Among these applications, NetLogo [42] provides an environment to rapidly under-
stand the value of ABM Digital Twin integration because it:
• Supports the FIPA and ACL standards through extensions;
• There is built-in support for connecting GIS data;
• Is being discussed in the Digital Twin Consortium, Mobility and Transportation
Working Group, to implement an air traffic use case used to interact with different
types of agents (e.g., positioner, evaluator, service provider, and user agents) and
has been used for similar applications [30], and;
• Has human-like (Belief, Desire, and Intention) and communication extension code
that could conform to the agent specification [43] by replacing existing NetLogo
extensions.

NetLogo provides a conceptual and theoretical overview providing a powerful

tool for the initial prototype. It is believed that the models should be simulator-
independent. Connections to the metaverse virtual reality, evident in the example of
the John Wayne Airport (JWA) terminal of the airport planning model that follows,
will enable the usage of realistic airport maps and data. Integrating real and near-time
data, along with the 3D reality capture imagery and the Digital Twin graph interface,
will enable distributed multi-agent modeling and simulation architectures to propa-
gate through the metaverse visualizations, modify existing visualizations, generate
synthetic 3D data scenes, and offer theoretical and empirical metaverse versions of
the predictive future simulation results. The goal of the agent simulation to the air-
port SoS is to provide a degree, magnitude, or measure of proposed types of agent
satisfaction by performing parameter sensitivity analysis in a scalable, independent
manner. Our agent approach will aid an eventual split of monolithic systems into
multi-agent networks, simplifying the airport’s overall architecture and improving
controllability and reliability since control functions can become autonomous and
failing agents can be restarted. The next section provides a guideline for applying
such models in airport planning.

2 Airport Planning Model

This section provides the workflow of airport planning and operational models. These
workflows will create a basis for the smart airport Agent-based Modeling design and
implementation explained in Sect. 3.
Distributed AI Modeling and Simulation for Smart Airport … 205

Table 2 Examples of environmental scenarios

Environment
Range Climate Social
Gold Severe Weather (below .−10.◦ ) Severe Disruption
Silver Poor visibility (.−10.◦ to 0.◦ ) Air Traffic Strike
Bronze Medium visibility (0.◦ –10.◦ ) Train Strike, Olympics
Normal Normal visibility (10.◦ –27.◦ ) Normal

2.1 Context-Planning Capabilities

Many parties are involved in generating schedules for all activities across the airport.
The multiple interdependencies between schedules mean that it may take weeks or
months to develop a default schedule. The business rules that a schedule may need
to take into account are not always written down, making it more difficult and time-
consuming to generate the schedule and to verify it against operational scenarios.
Airports operate within an ecosystem, with multiple parties providing and con-
suming services within the ecosystem. Whether a hub or point-to-point, an airport
acts as an intermodal transportation hub supporting a wide variety of transport modes,
which will increasingly consist of unmanned vehicles subject to automated control.
Interoperability in the SoS depends on the effective exchange of information based
on a common understanding (human and machine) between the parties, including
the airport, airlines, air traffic management, and a multitude of other participants in
the ecosystem. Common ontologies and associated data standards are fundamental
to meaningful information exchange.
Consequently, airport planning is not conducted in isolation. The airport operator
provides information on demand and capacity to airlines. Control authorities such
as air traffic management, and local aviation authorities impose constraints such as
the minimum time between landing two aircraft of type F. Each organization has its
own capabilities, including landing aircraft, refueling aircraft, immigration, check-in,
baggage handling, airfield maintenance, etc. Other factors, such as climate conditions
(visibility level) and social events (see example in Table 2), will increase the level of
complexity in the operations.
The capabilities to do work require resources such as aircraft and surface transport
vehicles, terminal facilities, airport equipment and facilities (runways, taxiways,
stands, etc.), and ground handlers. These resources not only need to be deployed but
they must also align with many other organizations. Planning activity occurs in many
different places in many different organizations and the interdependencies must link
for the operation to work effectively.
All work in an airport operation occurs in an environment, usually consisting
of climatic conditions such as a working temperature range, and social conditions
such as train driver strike. Plans are created for an environmental scenario category.
Environment scenarios can change quickly.
206 K. Alexandridis et al.

Fig. 1 Plan management cycle and stages

2.2 Planning Cycle

All organizations involved in an airport operation have their planning cycle typified
by the following diagram in Fig. 1. Four stages of the plan management cycle are
demand and capacity management, work scheduling, work execution, and perfor-
mance management. The ideal situation is a smooth flow of information across these
four stages.
This high-level perspective helps develop planning activities’ processes and infor-
mation requirements. Considering the power of Digital Twin for real-time data
exchange and analytics and Metaverse for cross-collaboration in a virtual space, it
is crucial to understand the business process of airport operation, which is explained
here. Default schedules are created for all work activities based on a load-balanced
capacity profile in a specified environmental scenario, taking account of resource
usage constraints. The planner obtains the demand for airport resources from demand
and capacity management, including the scenario it operates within, for example, a
normal day. This constitutes the required capacity profile. Demand and capacity
management also provide the resource usage constraints for the scenario.
Distributed AI Modeling and Simulation for Smart Airport … 207

Fig. 2 Workflow schedule standard model

The planner defines the business rules for the airport, airline, and government
work units within the airport for the provided scenario, including the default work
dependencies, the default work criteria, and the default resources to execute the work
(for example, in Fig. 2).
All schedules are verified against forecast assumptions, usually via a manual
process or automatically by process simulation or flow emulation. However, real-time
data is lacking in most of the cases. The outcome is a set of work units that do or do not
meet performance targets. Once default schedules have been verified and activated,
they typically need to be adjusted on the day to take account of circumstances such
as late arriving or departing flights, bad weather, and staff absence.
Given the lack of real-time data analytics and a single point of truth for stakeholder
collaborations (Metaverse) changes to plans may take many hours, particularly if
they have to be modified manually. Resources such as gates, stands, and baggage
reclaims are unlikely to be allocated efficiently by time-consuming manual planning
processes. This results in delays for flights, delays for passengers, and increased
airport costs.
On the effective date and time of the schedule, party role, transport, equipment,
and facility resources are allocated to the work. A system that automatically transfers
and integrates on-time data to automate the plan update is critical. Automating the
planning processes would reduce the time required to assess the impact of changes
from hours to minutes. The corresponding impact would be reduced delays, reduced
queues, and reduced costs.
208 K. Alexandridis et al.

Integrating data between applications can be expensive to develop and challeng-

ing to maintain. Standards for defining and exchanging information greatly reduce
the time, effort, and cost involved. A plan’s standard model improves the current
exchange of information to support multiple stakeholder requirements and is exten-
sible to incorporate future requirements. This challenge indicates the critical role
of Digital Twins for on-time/real-time data integration, analysis, and scheduling
updates. Furthermore, given the high complex and multi-stakeholder nature of air-
ports’ planning cycle, a multi-agent-based approach designed to handle the interac-
tion of stakeholders (as agents) will enrich the Digital Twin’s efficiency. To achieve
this goal, leveraging standard data schema is crucial. The next section provides an
overview of enriching The Aviation Community Recommendation Information Ser-
vices (ACRIS) data and information exchange standards by agent-based modeling
technologies.

2.3 Operational Level Integration

Planning closely relates to the operational and functional cycles at an airport. Such
operational levels are tied to the airport’s physical and virtual or computational infras-
tructure functions. Figure 3 provides an overview of the proposed ACRIS Semantic
Model [44] integration with the coupled Digital Twin and agent-based modeling
technologies. The three basic groups of category package modules of the seman-
tic model (namely entities, moments, and motivations) are closely related to agent
classes (as entities), to agent actions (as moments), and agent belief-desire-intention
(BDI) goal architecture [45, 46] (as motivations). Therefore, according to Fig. 3, a
Digital Twin for a Smart Airport can be enriched by the standard semantic models
(ACRIS) coupled with agents’ behavior to realistically reflect the airport’s operation.
More specifically, the knowledge organization of the ACRIS Semantic Model
integrates a number of operational modules, packages and data elements at the air-
port operational level [47] as can be seen in the following Fig. 4. Each of the three
fundamental objects (entities, moments, motivations), is mapped to each of the ele-
ment object library modules shown on the right side of Fig. 4. Each of these libraries
in turn has a number of nested attributes, operational layers and data associated
with them, and can be linked to the resources available to multi-agent BDI goals,
strategies, decisions, and actions.
At the visual, 3D, and metaverse levels of reality capture, these operational ele-
ments of the DT-ABM-ACRIS integrative architecture can be visually, geospatially,
and locationally oriented in virtual metaverse space. The following images in Fig. 5
demonstrate an example of the reality capture process of the John Wayne Airport
(JWA) terminal in Orange County, California. The combination of stationary, mobile,
and aerial LiDaR scanning technologies allows us to fully capture and closely asso-
ciate centimeter-accuracy locational attributes of captured scenes, 3D objects and
point-cloud coordinates with the developmental aspects of the agent-based model-
ing and simulation process.
Distributed AI Modeling and Simulation for Smart Airport … 209

Fig. 3 Digital Twin and agent-based model integration with basic module package organization of
the ACRIS semantic model

2.4 Definitions

2.4.1 Allocation

Allocations are a manifestation of the schedule in the current time. Allocation

refers to the assignment of resources to do work according to a published sched-
ule. For example, a schedule of aircraft landing slots will identify the runway facility
assigned to each particular landing slot. The runway facility will be allocated to
the landing at the scheduled time and unallocated to the landing when it is com-
plete.

2.4.2 Business Rule

Business rules are used to define the work to be done, the default work criteria, the
default resource types that are required to do the work, and the dependencies between
the work units. For example, the work unit “Land Aircraft” applies to all aircraft types
within the International Air Transport Association (IATA) aircraft type codes and will
require the default runway facility resource and also the participation of the air traffic
authority party role resource. The work unit “Taxi Aircraft to Stand” depends on the
210 K. Alexandridis et al.

Fig. 4 Package module and design organization of the modified ACRIS semantic model

“Land Aircraft” work unit having been successfully completed. This will apply to all
aircraft types and will require a taxiway and air traffic authority as default resources
(see example in Fig. 6).

2.4.3 Resource Criteria

Default Resource Criteria describe what a party wants to use resources for, such as
equipment or facility resources. Resource criteria define what a particular facility or
resource is required to be used for.

2.4.4 Resource Usage Constraint and Criteria

Resource constraints (usage constraints) specify the kind of work that a resource
cannot be used, for example, where an explosive detection system cannot be used
for Level 3 hold baggage screening. This is the opposite of resource usage criteria.
Distributed AI Modeling and Simulation for Smart Airport … 211

Fig. 5 Infrastructure operational visualizations of the John Wayne Airport (JWA) in Orange County,
California. Top Row: a, b LiDAR point-cloud and 360.◦ photosphere reality capture; c geodetic
monument examples at and around John Wayne Airport; d JWA monuments and high spatial
accuracy geodetic control network; e Orange County GPS real-time network ties

Fig. 6 Example inheritance and dependency structure of airport operation business rules
212 K. Alexandridis et al.

Table 3 Examples of integrated resource usage criteria

Type.a Resource.b Usage.c Criteria.d Example N.e B.e S.e G.e
F R-27L A F A380 1.0 0.5 0.2 0.1
F R-27L A F 747 2.0 1.0 0.5 0.2
F R-27L A D 3.0 2.0 1.0 0.5
F R-27L A C 4.0 3.0 2.0 1.0
F R-27L A B 4.0 3.0 2.0 1.0
F R-27L A A 5.0 4.0 3.0 1.0
E CD B SB 0.5 0.2 0.1 0.01
E CD B OOG 0.0 0.0 0.0 0.0
E OOGD B SB 0.1 0.1 0.1 0.01
E CBX P PS 6.0 3.0 2.0 1.0
E RCAR B SB 20.0 15.0 10.0 2.0
a
. F: Facility; E: Equipment
b
. R-#: Runway Number; CD: Check Desk; OOGD: OOG Desk; CBX: Cabin Bag X-Ray; RCAR:
Reclaim Carousel
c
. A: Aircraft; B: Bag; P: Person
d
. A-F: Code; SB: Standard Bag; OOG: OOG Bag; PS: Passenger
e
. Operation cycles per minute: N-Normal, B-Bronze, S-Silver, G-Gold

Resource Usage Criteria, or just Usage Criteria, define what a particular facility
or resource can be used for. This is the opposite of resource (usage) constraints.
Resource usage criteria implicitly provide constraints on resources, for example
standard check-in desks can deal with standard bags, but not Out Of Gauge (OOG)
bags. An OOG desk can deal with standard bags as well as OOG bags, though at a
reduced throughput in terms of bags per minute.
Usage criteria and capacities are provided to the planner by demand and capac-
ity management and they may differ for each environmental scenario category. An
example of a set of resource usage criteria and their characteristic properties is shown
in Table 3.

2.4.5 Work Criteria and Progress Workflow

Work is a combination of work criteria and the resource that is to be applied to the
work. It typically comprises a hierarchy of units of work. For example the work to
land an aircraft includes taxiing the aircraft to the stand and parking it at the stand.
Work may have dependencies, e.g., aircraft separation must be maintained to 5 min
for code F, and may have a specified order in which it needs to be done, e.g., unload
baggage from aircraft and then move Unit Load Devices (ULDs) to arrivals break.
Work Criteria is a standard for defining work to be done, including sufficient
information so that the work can be:
• Scheduled with competent resources.
• Readily determined to have been completed or not.
• Objectively measured.
Distributed AI Modeling and Simulation for Smart Airport … 213

Work criteria identify the type of work to be done, e.g., (1) unload baggage from a
British Airways 744 aircraft at 08:00 every Monday during the 2023 summer season,
or (2) sort and load bags for business class passengers traveling on a British Airways
flight to Boston. See Fig. 7 for further examples in the context of the required work.
Work progress is the quantifiable completion of work. The actual doing of the work
is part of logistics and conformance. Monitoring the work progress by resources and
assets is part of performance management.
In the plan model, a plan is a forward-looking work schedule. A work schedule
is an association of facility, equipment, and party role resources to units of work in
specified time periods. It takes business requirements into account, as provided in a
load-balanced capacity profile for a defined scenario. It also incorporates the usage
constraints of the resources that are to be included in the work schedule. Units of work
should have a global unique identifier to enable accurate allocation and deallocation/
reallocation of resources to the work.

3 Smart Airport ABM Design and Implementation

3.1 General Considerations

As we indicated, the time-consuming and manual resource planning and alloca-

tions, such as gate stands and baggage reclaims, are highly challenging and often
generate delays for flights, delays for passengers, and increased airport costs. The
work schedule and sequence of activities can be designed, monitored, and controlled
through agent-based modeling and simulation approaches implemented on a Digital
Twin infrastructure. For instance, [14] examined the passengers’ (as agents) satisfac-
tion and the time-saving options through their movements toward different services
such as passport controls, check-in counters of airline companies, boarding gates,
and different types of shopping. This reflects the passengers’ behavior, discretionary
decisions (shopping), and the number of services (passport controls and boarding
gates) to examine different scenarios.
The results of the 30-round simulation indicate that satisfaction decreases as
service lines increase and user agents miss their flights. The increasing number
of service lines can be due to delays in passport checks or limited airline check-in
services.
Furthermore, the simulation results demonstrate that average time lengthens when
a very high number of agents are included in the simulation. Reflecting the lack of
optimized scheduling for 6216 incoming or outgoing passengers to/from the termi-
nals.
Therefore, using two evaluation criteria (time savings and agent satisfaction) the
potential benefits of using ABM are highlighted. However, it should be noted that
this work is a proof of concept and NetLogo simplifies the process and reduces
the number of agents to address the computing requirements. As such, the focus of
214 K. Alexandridis et al.

Fig. 7 Arrival time probability distribution segmentation for incoming (left subgraph) and outgoing
(right subgraph) passengers of a single flight in an airport terminal

NetLogo ABM is on a limited number of options. Therefore, further investigation

and more robust ABM tools should be utilized to examine the ABM application in
a Digital Twin.
Capturing passenger flow experiences related to travel times within, to, and from
airports requires a firm understanding of segmentation across discrete and distinct
spatiotemporal sections within any given airport terminal. Relevant literature has
considered longitudinal methods of travel modes (e.g., [48]) in a smart city environ-
ment and statistically estimated wait and travel time distributions for citizens and
travelers. Rather than modeling and simulating random passenger arrival times, we
plan to use well-known Poisson probability distributions (see, for example, [49, 50]
on how these distributions relate to human and socio-temporal dynamics). The fol-
lowing Fig. 7 showcases temporal distribution dynamics based on these parametric
probabilistic inferences for both incoming and outgoing single flight passengers by
airport experience segment.
While day-to-day operational reality in a single airport is by far more complex
than the oversimplification shown in the figure (e.g., hundreds of inbound and out-
bound flights per day, many of them overlapping across the temporal continuum),
a distributed agent-based simulation framework can combine these arrival proba-
bilities based on daily airport schedule of flight arrival and departures and generate
evidence-based, empirical mixture (e.g., multimodal) distributions to be used as
Bayesian Priors for the BDI decision-making architecture.

3.2 Agent-Based Architecture

3.2.1 Smart Airport Agent Typology

A number of conceptual elements of a smart airport agent-based architecture can be

considered as fundamental elements of an agent typology. Employing a distributed
modeling architecture requires the differentiation of agent archetypes. We can iden-
tify at least four fundamental types of agents, as illustrated in Fig. 8:
Distributed AI Modeling and Simulation for Smart Airport … 215

Fig. 8 An example of smart airport agent typology considerations

• Passenger-type functional agents.

• Flight-related operational agents.
• Terminal-related operational agents.
• Vendor and concession-related functional agents.

Adopting such a distributed agent typology schema enables the introduction and
use of an agent orchestration framework within a broader BDI agent architecture
[45, 51]. Specifically,
Passenger-Agent: This archetypal agent typology captures and monitors basic
passenger-agent spatial movements across different airport and terminal sections.
It begins with the passenger transportation mode section by tracking the mode of
transportation (e.g., public transport or service transportation) along with parking
and parking spot allocation. Once passengers move into the main airport terminal
(e.g., airline check-in section and location), it tracks flight assignments, thus associ-
ating each passenger agent with a specific flight module (consequently, linking the
passenger-agent module with the flight-agent and terminal-agent modules). Further,
along the passenger journey through the airport, the passenger-agents relate their
location and assignment through the security check-in counters/locations, the choice
of vendors, and vendor location for airport concessions (thus associating passenger-
agent entities with vendor-agent entities). Finally, it tracks passenger movement
through their departure/arrival gates and embarkation/disembarkation queues.
Flight-Agent: This typology provides a distributed agent-based framework for all
flight-related components or modules related to smart airport operations. These
include the check-in management process (from the airline and provider operational
perspective), gate assignment for flights departing or arriving, time and process man-
agement for departures and arrivals, including delay and queue management, plane
movement within an airport (e.g., from/to gates, etc.), plane turnaround and airport
collaborative decision-making processes (A-CDM), as well a framework for modu-
larity interfacing agent roles with airline operational systems.
Terminal-Agent: The third typology relates to smart airport terminal-based oper-
ations. These agents track and distribute baggage movements, gate management
assignments (from an airport facility operational perspective), facility maintenance
and integrated management (in terms of both temporal and spatial perspectives), air-
216 K. Alexandridis et al.

port security (physical and network), and smart asset management. These terminal-
agent types interface directly with relevant system modules and provide distributed
intelligent communication interfaces with physical system modules.
Vendor-Agent: This agent typology relates to distributed communication and action
mechanisms related to vendor assignments, operations, product and service delivery
mechanisms, concession property and title management, and interface (externally)
with physical and virtual point-of-sale interfaces within an airport smart operations
system.
Collectively, these examples of smart airport agent typologies are capable of pro-
viding the necessary role assignments and functions that enable a smart, robust, adap-
tive, and intelligent framework for agent actions, resource (management/allocation/
distribution), time and space interactions with physical movement, behavioral and
cognitive response mechanisms, and property and facility management functions.
While not necessarily exhaustive (in fact, alternative and augmented additional agent
types and typology groups can be designed and used), they form a cohesive and tightly
coupled set of agent roles that, through their exchanges and interactions, can give
rise to the emergence of complex and intelligent distributed system behaviors and
operational mechanisms.

4 Implementation in a Metaverse Visual Framework

This section focuses on the visual metaverse aspects of an intelligent Digital Twin
framework. It is mainly based on an award-winning TM Forum Moonshot Catalyst
project (URN M23.0.567) [13]. The project features innovative and active public-
private partnerships across a range of industry consortia, public organizations, and
airports. The remaining section content provides a short and concise outline of the
reference architecture and design as well as the structural and operational intelligence
modeling and simulation elements used in enhancing a Digital Twin coupled with
metaverse experience, supporting the implementation of the agent-based intelligent
ontology semantic planning framework described in the previous sections.
Figure 9 provides an overview of the key elements of the interoperable airport
ecosystem for a metaverse approach. The elements range from those of high abstrac-
tion/fundamental composability (e.g., public transport, airports, airlines, passengers,
agencies), to more specific and metaverse domain-explicit concepts (retail, telecom,
AECO, traffic management, etc.). The intent-based service composition and orches-
tration, along with information exchange and cross-domain interoperability, are gov-
erned by a common and shared ontology framework, such as the ACRIS semantic
framework described in Sect. 3.
The relevant reference implementation outline example is shown in Fig. 10. As the
back-end metaverse framework interacts with the entities, moments, and motivators
(e.g., person recognition and hazard recognition service layers), the mid-end layers
provide a design and reference orchestration for services like application communi-
cation, scenario modeling and simulation, data transformation and interoperability,
Distributed AI Modeling and Simulation for Smart Airport … 217

Fig. 9 Elements of the interoperable airport ecosystem metaverse approach. The scalable archi-
tecture includes and encapsulates a common semantic ontology for cross-domain interoperability

Fig. 10 Real-time information exchange components and layers for metaverse reference imple-
mentation for the Digital Twin smart airport model

and validation and assessment services. Both network-centric and knowledge-centric

orchestration and coordination service layers operate across a range of coordinated
layers to achieve integrative delivery and visualization services. The latter (3D visu-
alization services) allow for metaverse immersive and interactive experiences while
enabling and facilitating the provisioning of other end-user (front-end) virtualization
218 K. Alexandridis et al.

Fig. 11 Metaverse visualization and real-time Digital Twin dashboard operations: a a visual display
of navigational agent spatial movements; b visual check-point operations and service message
exchange real-time operations; c showcase of interactive airport parking allocation and navigation;
d passenger-based dashboard for real-time event-driven scenario simulations

services, such as travel management and product fulfillment-all the service orches-
tration and coordination/orchestration operations function within a near-real-time,
high-fidelity functional metaverse environment.
In terms of the metaverse visualization experience (see Fig. 11), the nature and
characteristics of Digital Twin 3D immersion allow for both enhancing the visual
intelligence and locational/spatial navigational experience and the near real-time ser-
vice, data, and communication provision. While the former enhances and supports
visual intelligence operations, the latter promotes and encapsulates implementation
based on high accuracy, fidelity, and interoperability. Taken together, these two fun-
damental principles (visual intelligence and fidelity), when applied within a Digital
Twin metaverse model, work complementary to each other in establishing and bench-
marking a conceptual framework for intelligent Digital Twins.
The overview of the overall smart airport and intelligent Digital Twin metaverse
implementation is provided in Fig. 12. The approach focuses on predictable, accurate,
and visually immersed outcomes stemming from automation, intelligent decision-
making, and inter-agent distributed intelligence. Both internally (within events, ser-
vice provisioning, resource allocation components, and orchestration mechanisms)
and externally (inputs and outputs such as data and services), it is essential to strive
to achieve a high degree of cross-standards interoperability and composability.
Some of the benefits of integrative metaverse design, control, and operations
include the ability of airport management and operational planning capable to sim-
ulate and experimenting with alternative scenarios before implementation, reducing
Distributed AI Modeling and Simulation for Smart Airport … 219

Fig. 12 Visual conceptualization of the intelligence airport metaverse framework

cost (design and implementation) while managing risk and risk expectations (includ-
ing risk aversion and risk propensity attitudinal dispositions). As a result, an intelli-
gent metaverse Digital Twin framework streamlines time-to-value operations while,
in the near- and long-term, promoting higher return-on-investment ratios. At the
same time, the Digital Twin framework enables the development and establishment
of broader sustainability and robustness characteristics at the airport level. These
might include emergency and incident response or reporting operations, addressing
sustainability goals and criteria, and forming a more innovation-minded workforce.

5 Conclusions and Future Work

In this chapter, we proposed a new generation of ABM that leverages the metaverse
and Digital Twin capabilities. Also, we highlighted the significant role of real-time
data for a more realistic representation of Airport operational activities and planning
processes. The aim is to optimize and enable airport planning with predictive analysis
to understand and adapt to unpredictable issues. We argued that the complexity of
airport systems cannot be addressed by linear and ad hoc modeling and simulation
methods. To address this challenge, we described two methods of categorizing ABMs
to set up both the use of phenomena and exploratory-based models in the airport
planning domain and introduce agent and communication standards. It was shown
220 K. Alexandridis et al.

that many agent systems use the standard Foundation for Intelligent Physical Agents
and Agent Communication Language (FIPA-ACL) to support the sharing of message
content with common semantics that includes communication context. This context
explicitly defined relationships between a particular message in relation to a context
such as a sender agent’s current workflow, an agreed interaction protocol, goals, and
plans, or with respect to a receiving agent’s beliefs, desires, and intentions [53]. The
OMG Agent and Event Metamodel [31] standard will supersede FIPA-ACL but is
planning to include ACL compliance. Future work will trade-off ACL working with
the message exchange formats and transport/application protocols.
We provided a key perspective to airport planning model development, including
an ecosystem for context planning, a planning cycle with a workflow standard model,
integration of the Digital Twin, metaverse, and ABM model, and key definitions.
The provided conceptual model has practical and theoretical implications. From a
practical perspective, different operational activities can benefit the role of Metaverse
and Digital Twins to test the human-environment interactions in various scenarios.
This capability can be used for the interaction of different stakeholders, such as air-
lines, third-party services, and safety and security teams. The agent-based modeling
and simulation, calibrated with real-time data, is a new line of research that needs
further exploration to offer a reliable and trustworthy output for decision-makers
in airport operations. In addition, an ontological framework is critical to address
the interoperability and harmonization of complex data and analytics in an airport
ecosystem.
Finally, this document provided an agenda for developing a proof of concept
and pilot study to be implemented in an actual case and an examination of the
idea of a new generation of ABM integrated into the Digital Twin architecture and
detailed information and guidelines for Airport Planning, which can be used for
ABM development.

Acknowledgements The work presented in this chapter draws from a fertile collaborative environ-
ment for discussions, discourse, and collaboration of a number of collective organizations. Digital
Twin Consortium (DTC) working groups are one of these organizations. S. Sabri and K. Alexan-
dridis serve as co-chairs of the DTC Academia and Research Working Group, while D. Migliori
serves as one of the co-chairs of the DTC Mobility and Transportation Working Group. Many of
these ideas and concepts emerged through its collaborative environment and the author’s interac-
tions within it. The parent industry standards development organization Object Management Group
(OMG) is another key organization we wish to acknowledge. Jeff Smith serves as co-chair of the
OMG Analysis and Design Task Force and co-submitter of the AgEnt specification. TM Forum and
the Airports Council International (ACI) provided the necessary background for the development
of ideas, projects, and resources. Specifically, the metaverse implementation related to passenger
experiences, provided in Sect. 4, is an integral part of an award-winning Moonshot Catalyst program
of TM Forum and represents an active collaboration of many industry and public organizations.
D. Migliori, B. Logan, and K. Alexandridis serve as key organizational and industry champion
members of the project. Furthermore, the integration of the ACRIS Semantic Model described in
Sect. 3 owes its roots and fundamental concepts to the ACI’s Aviation Community Recommended
Information Services (ACRIS) model approach. B. Logan is one of the key architects and custo-
dians of the ACRIS semantic model organization and its resources. Finally, the authors wish to
acknowledge a number of individuals and organizational representatives for the exchange of ideas
and organizational support: Dan Isaacs, general manager of OMG and CTO of DTC for his sup-
Distributed AI Modeling and Simulation for Smart Airport … 221

port of DTC working groups and organizational facilitation; Members of various DTC working
groups for feedback and review of this work and ideas over various stages; Member-organizations
of the TM Forum Moonshot Catalyst Project No. M23.0.567 for supporting and recognizing the
transformative concepts.

References

1. Mekić A, Mohammadi Ziabari SS, Sharpanskykh A (2021) Systemic agent-based modeling

and analysis of passenger discretionary activities in airport terminals. In: Aerospace, vol 8, p
162, June 2021. 3 citations (Crossref) [2023-12-04] Number: 6 Publisher: Multidisciplinary
Digital Publishing Institute
2. Cheng L, Reddy V, Fookes C, Yarlagadda P (2014) Agent-based modelling simulation case
study: assessment of airport check-in and evacuation process by considering group travel
behaviour of air passengers, vol 568–570, pp 1859–1864, Trans Tech Publications Ltd, 2014.
4 citations (Crossref) [2023-09-25]
3. Zhao C, Dai X, Lv Y, Niu J, Lin Y (2023) Decentralized autonomous operations and organi-
zations in transVerse: federated intelligence for smart mobility. IEEE Trans Syst Man Cybern
Syst 53(4):2062–2072
4. El Jaouhari A, Arif J, Samadhiya A, Kumar A, Jain V, Agrawal R (2023) Are metaverse
applications in quality 4.0 enablers of manufacturing resiliency? An exploratory review under
disruption impressions and future research. TQM J
5. Kleinschmidt T, Guo X, Ma W, Yarlagadda PK (2011) Including airport duty-free shopping in
arriving passenger simulation and the opportunities this presents. In: Proceedings of the 2011
Winter Simulation Conference (WSC), (Phoenix, AZ, USA), pp 210–221, IEEE, Dec 2011. 2
citations (Crossref) [2023-09-25]
6. Tesoriere G, Campisi T, Canale A, Severino A, Arena F (2018) Modelling and simulation of
passenger flow distribution at terminal of Catania airport. In Proceedings of the international
conference of computational methods in sciences and engineering 2018 (ICCMSE 2018),
(Thessaloniki, Greece), p 140006. 17 citations (Crossref) [2023-09-25]
7. Ma W, Fookes C, Kleinschmidt T, Yarlagadda P (2012) Modelling passengers flow at airport
terminals—individual agent decision model for stochastic passenger behaviour. In: Pina N,
Kacprzyk J (eds) Proceedings of the 2nd international conference on simulation and modeling
methodologies, technologies and applications (SIMULTECH). SciTePress Digital Library, pp
109–113
8. Helbing D, Buzna L, Johansson A, Werner T (2005) Self-organized pedestrian crowd dynamics:
experiments, simulations, and design solutions. Transp Sci 39:1–24. 1019 citations (Crossref)
[2023-09-25]
9. Doostmohammadi M, Fragniere E, Holdsworth R (2020) Crowd behaviour modelling devel-
opments through mixed integer programming: the case of Airport queue management. In:2020
9th International Conference on Industrial Technology and Management (ICITM), pp 32–36
10. Bandarian F, Kordani A (2022) Evaluation of passenger behavior in the baggage claim area of
the Airport passenger terminal. In: Mathematical problems in engineering, vol 2022. 0 citations
(Crossref) [2023-09-25]. Hindawi Limited
11. Zhu H, Liu S, Li X, Zhang W, Osgood N, Jia P (2023) Using a hybrid simulation model to
assess the impacts of combined COVID-19 containment measures in a high-speed train station.
J Simul. 2 citations (Crossref) [2023-09-25] Taylor and Francis Ltd
12. Liu X, Fu H (2023) Simulation and optimization of inbound passenger flow line in large-scale
railway station. In: Kong X, Falcone F (eds) Proceedings of SPIE—the international society
for optical engineering, vol 12708, SPIE. 0 citations (Crossref) [2023-09-25]
222 K. Alexandridis et al.

13. TM Forum (2023) Transforming passenger experiences with continuous decision intelligence
(Moonshot Catalyst URN M23.0.567)
14. Zhang H, He B, Lu G, Zhu Y (2022) A simulation and machine learning based optimization
method for integrated pedestrian facilities planning and staff assignment problem in the multi-
mode rail transit transfer station. In: Simulation modelling practice and theory, vol 115. 9
citations (Crossref) [2023-09-25] Elsevier B.V
15. Peng J, Wei Z, Wang S, Qiu S (2023) Toward dynamic regulation of bidirectional automatic fare
gates: a macro-to-micro simulation approach. In: Simulation modelling practice and theory,
vol 124. 1 citations (Crossref) [2023-09-25] Elsevier B.V
16. Delcea C, Cotfas LA, Milne R, Xie N, Mierzwiak R (2022) Grey clustering of the variations in
the back-to-front airplane boarding method considering COVID-19 flying restrictions. In: Grey
systems, vol 12, no 1, pp 25–59. 8 citations (Crossref) [2023-09-25] Emerald Group Holdings
Ltd
17. Zellers SJH, Bradford WC, Snowden CH, Knoedler RJ, Alexander JE, Coalson MB, Sonkin
J, Foley C, Stinson W, Tansey P, Smith JF, Quinn J, Williams K, Whitfield-Smith L (2024)
Airport energy resiliency roadmap. Transportation Research Board, Washington, DC
18. Xianliang G, Jingchao X, Zhiwen L, Jiaping L (2021) Analysis to energy consumption char-
acteristics and influencing factors of terminal building based on airport operating data. In:
Sustainable energy technologies and assessments, vol 44. 4 citations (Crossref) [2023-09-25]
Elsevier Ltd
19. Gu X, Xie J, Huang C, Ma K, Liu J (2022) Prediction of the spatiotemporal passenger distribu-
tion of a large airport terminal and its impact on energy simulation. In: Sustainable cities and
society, vol 78, p 103619. 5 citations (Crossref) [2023-09-25]
20. Gu X, Xie J, Huang C, Liu J (2022) A spatiotemporal passenger distribution model for airport
terminal energy simulation. In: Indoor and built environment, vol 31, pp 1834–1857. 1 citations
(Crossref) [2023-09-25]
21. Li Z, Zhang J, Mu S (2023) Passenger spatiotemporal distribution prediction in airport terminals
based on insect intelligent building architecture and its contribution to fresh air energy saving.
In: Sustainable cities and society, vol 97, p 104772. 0 citations (Crossref) [2023-09-25]
22. Lin L, Liu X, Liu X, Zhang T, Cao Y (2023) A prediction model to forecast passenger flow
based on flight arrangement in airport terminals. In: Energy and built environment, vol 4, no
6, pp 680–688. 1 citations (Crossref) [2023-09-25] KeAi Communications Co
23. Janssen S, Sharpanskykh A, Curran R (2019) Agent-based modelling and analysis of security
and efficiency in airport terminals. In: Transportation research part C: emerging technologies,
vol 100, pp 142–160. 18 citations (Crossref) [2023-12-04]
24. Zhang X, Chen G, Shi J (203) Simulation of guided crowd evacuation scheme of high-speed
train carriage. Int J Simul Modell 22(1):110–120. 0 citations (Crossref) [2023-09-25] DAAAM
International Vienna
25. Lahoti N (2022) How the metaverse will change transportation and future mobility
26. Darvishi M, Ahmadi G (2014) Validation techniques of agent based modelling for geospatial
simulations. Int Arch Photogrammetry Remote Sens Spatial Inf Sci XL-2-W3:91–95. 8 citations
(Crossref) [2023-12-04] Conference Name: WG II/1, WG II/4, ICWG II/IV, WG IV/7, The
1st ISPRS International Conference on Geospatial Information Research (Volume XL-2/W3)
- 15–17 November 2014, Tehran, Iran Publisher: Copernicus GmbH
27. Smaldino P (2023) Modeling social behavior: mathematical and agent-based models of
social dynamics and cultural evolution. Princeton University Press. Google-Books-ID:
aOS3EAAAQBAJ
28. Scala P, Mota M, Blasco-Puyuelo J, Garcia-Cantu O, Blasco C (2022) A novel validation
approach for validating the simulation model of a passengers’ airport terminal: case study Palma
de Mallorca airport. In: European modeling and simulation symposium, EMSS. 0 citations
(Crossref) [2023-09-25]
29. Wilensky U, Rand W (2015) An introduction to agent-based modeling: modeling natural,
social, and engineered complex systems with NetLogo. MIT Press, Google-Books-ID, LQrhB-
wAAQBAJ
Distributed AI Modeling and Simulation for Smart Airport … 223

30. Carbo J, Sanchez-Pi N, Molina JM (2018) Agent-based simulation with NetLogo to evaluate
ambient intelligence scenarios. J Simul 12:42–52. 9 citations (Crossref) [2023-09-25] Pub-
lisher: Taylor & Francis _eprint: https://doi.org/10.1057/jos.2016.10 Northwestern: https://
ccl.northwestern.edu/2016/carbo.pdf
31. OMG (2022) Agent and Event Metamodel (AgEnt) joint revised submission. "AgEnt" Team
0.4-01, Object Management Group
32. Vlăduţescu, Stănescu G (2023) Environmental sustainability of metaverse: perspectives from
Romanian developers. Sustainability (Switzerland) 15(15)
33. Khalaj O, Jamshidi M, Hassas P, Mašek B, Stadler C, Svoboda J (2023) Digital twinning of
a magnetic forging holder to enhance productivity for industry 4.0 and metaverse. Processes
11(6)
34. Ma Z (2023) Energy metaverse: the conceptual framework with a review of the state-of-the-art
methods and technologies. Energy Inf 6(1)
35. Negi P, Singh R, Gehlot A, Kathuria S, Thakur A, Gupta L, Abbas M (2023) Specific soft com-
puting strategies for the digitalization of infrastructure and its sustainability: a comprehensive
analysis. Arch Comput Methods Eng
36. Naderi H, Shojaei A (2023) Digital twinning of civil infrastructures: current state of model
architectures, interoperability solutions, and future prospects. Autom Constr 149
37. Pamucar D, Deveci M, Gokasar I, Tavana M, Köppen M (2022) A metaverse assessment
model for sustainable transportation using ordinal priority approach and Aczel-Alsina norms.
In: Technological forecasting and social change, vol 182
38. Deveci M, Pamucar D,Gokasar I, Martinez L, Köppen M, Pedrycz W (2024) Accelerating the
integration of the metaverse into urban transportation using fuzzy trigonometric based decision
making. In: Engineering applications of artificial intelligence, vol 127
39. Grupac M, Negoianu A-E (2023) Immersive extended reality and sensor-based object recogni-
tion technologies, socially-oriented location tracking and simulation modeling tools, and arti-
ficial vision and haptic augmented reality systems in the metaverse interactive environment.
Rev Contemporary Philos 22:226–243
40. Grieves M, Hua E (2024) Defining, exploring, and simulating the Digital Twin metaverses. In:
Digital Twins, simulation, and metaverse: driving efficiency and effectiveness in the physical
world through simulation in the virtual worlds, Forthcoming. Springer, Heidelberg
41. Xu Z, Bai Q, Shao Y, Hu A, Dong Z (2022) A review on passenger emergency evacuation from
multimodal transportation hubs. J Traffic Transp Eng (English Edition) 9:591–607. 2 citations
(Crossref) [2023-09-25]
42. Wooldridge M, Jennings NR (1995) Intelligent agents: theory and practice. In: The knowledge
engineering review, vol 10, pp 115–152. 3452 citations (Crossref) [2023-09-25] Cambridge
University Press
43. Benfield SS, Hendrickson J, Galanti D (2006) Making a strong business case for multiagent
technology. In: Proceedings of the fifth international joint conference on autonomous agents and
multiagent systems, (Hakodate Japan). ACM, pp 10–15. 21 citations (Crossref) [2023-09-25]
44. Airports Council International (2019) ACI-ACRIS semantic model of SOA, tech rep, Airports
Council International (ACI)
45. Grimm V, Berger U, Bastiansen F, Eliassen S, Ginot V, Giske J, Goss-Custard J, Grand T,
Heinz SK, Huse G, Huth A, Jepsen JU, Jorgensen C, Mooij WM, Muller B, Pe’er G, Piou
C, Railsback SF, Robbins AM, Robbins MM, Rossmanith E, Ruger N, Strand E, Souissi S,
Stillman RA, Vabo R, Visser U, DeAngelis DL (2006) A standard protocol for describing
individual-based and agent-based models. Ecol Modelling 198(1–2):115–126. 1896 citations
(Crossref) [2023-09-25]
46. Alexandridis KT (2006) Exploring complex dynamics in multi agent-based intelligent systems:
theoretical and experimental approaches using the Multi Agent-based Behavioral Economic
Landscape (MABEL) model. Ph.D. Dissertation, Purdue University, West Lafayette, Indiana,
May 2006. ISBN: 9780542864230 Publication Title: Forestry & Natural Resources Volume:
Doctor of Philosophy 3232142
224 K. Alexandridis et al.

47. Alayande S, Airports Council International (2017) ACRIS semantic model guide to the knowl-
edge layer. Technical Report 1.1, Airports Council International (ACI)
48. Tenkanen H, Toivonen T (2020) Longitudinal spatial dataset on travel times and distances by
different travel modes in Helsinki Region. Sci Data 7:77. 16 citations (Crossref) [2023-09-25]
Number: 1. Nature Publishing Group
49. Agresti A (2003) Introduction: distributions and inference for categorical data. In: Categorical
data analysis, Wiley Series in Probability and Statistics. John Wiley & Sons - Wiley Inter-
Science, 2nd edn, pp 1–35, Section: 1
50. Crane R, Sornette D (2008) Robust dynamic classes revealed by measuring the response func-
tion of a social system. In: Proceedings of the national academy of sciences, vol 105, pp
15649–15653. 535 citations (Crossref) [2023-09-25]
51. Wai Sk, WaiShiang C, Khairuddin MAB, Bujang YRB, Hidayat R, Paschal CH (2021)
Autonomous agents in 3D crowd simulation through BDI architecture. JOIV : Int J Inf Visu-
alization 5:1–7. 2 citations (Crossref) [2023-12-04] Number: 1
52. Palanca J, Rincon JA, Carrascosa C, Julian VJ, Terrasa A (2023) Flexible agent architecture:
mixing reactive and deliberative behaviors in SPADE. Electronics 12:659. 1 citations (Crossref)
[2023-12-04] Number: 3. Multidisciplinary Digital Publishing Institute
53. Poslad S (2007) Specifying protocols for multi-agent systems interaction. ACM Trans
Autonomous Adaptive Syst 2:15–es. 94 citations (Crossref) [2023-12-04]
The First Real-Time Digital Twin
of a Nuclear Reactor

John Darrington, Ben-oni E. Vainqueur, and Christopher Ritter

Abstract In this chapter we examine the first fully operation digital twin of a fissile
nuclear reactor. In 2023, Idaho National Laboratory (INL) partnered with Idaho State
University (ISU) to create and run a digital twin of their AGN-201 fissile nuclear
reactor. This reactor can produce up to 5 watts and has proven to be the perfect
testbed for discovering how best to implement a digital twin of a larger reactor
system. INL created an integrated, cloud-based digital twin capable of measuring
data from the reactor in near real-time and built the foundation for running operations
such as machine learning and analytics in near real-time in an offsite location. This
experiment is a key step in building a digital twin of a larger reactor system and has
helped highlight many potential pitfalls and problems that such an endeavor might
face. This experiment has also shown the great promise that a cloud-first approach
has when creating digital twins.

Keywords Red vs. blue · Surrogate models · Isolation forest · Nuclear

proliferation · Augmented reality · Mixed reality · Virtual reality · Real-time
capture · Data analysis · Azure cloud · Microsoft azure · Grownet · Rust · Digital
twin security · Aqueous seperations · Anomaly detection system · Physics
accessibility · Microsoft hololens

1 A Brief History of INL’s Nuclear and Digital Twin Work

The Idaho National Laboratory (INL), established post-World War II, has played a
significant role in the development of nuclear technology, and now Digital Twins.
Initially named the National Reactor Testing Station, it was the site where, in 1951,
the Experimental Breeder Reactor I (EBR-I) first generated electricity from nuclear
power.

J. Darrington (B) · B. E. Vainqueur · C. Ritter

Digital Engineering Battelle Energy Alliance, National Laboratory USA, Idaho Falls, Idaho, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 225
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_10
226 J. Darrington et al.

During the 1960s and 1970s, INL focused on researching advanced reactor designs
and improving nuclear safety. By the 1980s, the laboratory shifted its attention to
environmental management, particularly the remediation of nuclear waste.
In recent decades, INL has continued to work on nuclear reactor technology,
emphasizing extending reactor lifespans, advancing next-generation reactors, and
contributing to small modular reactor (SMR) research. It also engages in cyber-
security for nuclear facilities and supports nuclear nonproliferation and security
initiatives.
Today, INL is involved in developing nuclear energy technologies with an
increased emphasis on addressing climate change. Throughout its history, INL has
had a substantial influence on the evolution of nuclear energy policy and infrastructure
in the United States.
For the past 3–5 years, the Idaho National Laboratory (INL) has been championing
the use and creation of DTs in the nuclear industry. This consists of developing virtual
DTs to evaluate potential reactor misuse, and for diversion scenarios for sodium fast
reactors and high-temperature Pebble Bed Reactors [1].
According to the United States Department of Defense, Digital Engineering is
a comprehensive approach that relies on authoritative sources of system data and
models, seamlessly spanning multiple disciplines to support activities throughout
a product’s entire lifecycle, from conceptualization to disposal [2]. Unlike tradi-
tional static documents, digital engineering emphasizes the use of interconnected and
dynamic models to describe a product. These models are integrated across diverse
platforms to facilitate the development of a design product that can be effectively
supported throughout its operational life [3, 4].
Digital Twins leverage the principles of digital engineering to create a model
that represents either an existing physical product or the design of a future physical
product [4]. This concept allows for the creation of two distinct types of Digital Twins:
a virtual Digital Twin, which simulates the behavior of a physical product through
modeling and simulation, and a Digital Twin of a physical system, which combines
a physical asset with computational models to provide a virtual representation of the
asset.

2 The Impact of Digital Twins in the Nuclear Sector

The adoption of Digital Twins has begun to revolutionize various aspects of nuclear
plant operations, maintenance, and training [1, 5, 6]. Plant reference simulators serve
as a rudimentary form of Digital Twins, providing operators with a training tool.
However, these simulators are not directly connected to the actual power plant.
Furthermore, Digital Twins are increasingly being applied in the realm of condi-
tion monitoring. Online performance and condition data can be collected through
sensors, which are then processed by various artificial intelligence and machine
learning algorithms to provide valuable insights.
The First Real-Time Digital Twin of a Nuclear Reactor 227

Training of operators has seen improvements using plant reference simulators,

which, despite acting as a form of Digital Twin, have their limitations due to the lack
of a real-time connection with the actual nuclear power plants. However, these simu-
lators are critical for providing operators with a realistic, yet controlled environment
to hone their skills without impacting plant operations.
Additionally, the integration of DTs is becoming increasingly prevalent within
the current nuclear fleet, marking a significant advance in enhancing operational
efficiency, maintenance practices, and training methodologies [1, 5, 6]. Digital Twins
serve as dynamic, virtual models of the nuclear plants and can significantly aid in
simulating plant processes, facilitating a proactive approach to plant management.
Furthermore, the application of DTs extends into the realm of condition moni-
toring. By employing an array of sensors, continuous streams of performance and
condition data are captured in real-time. This influx of data is not merely collected;
it is analyzed and leveraged by advanced AI/ML algorithms. The result is a highly
informed, data-driven approach to monitoring the plant’s health, which can predict
potential issues, optimize performance, and reduce downtime. By tapping into the
capabilities of AI and ML, the data harnessed through DTs is transformed into action-
able insights, allowing for smarter, more efficient plant management. This integration
points to a future where digital and operational technologies converge to create a more
reliable and safe nuclear power infrastructure.

2.1 Aqueous Separations

Another notable application of Digital Twins in development pertains to aqueous

metal separations. The Idaho National Laboratory (INL) is actively involved in the
creation of a Digital Twin that emulates an aqueous separation process utilizing
centrifugal contactors. This Digital Twin is expected to play a vital role in moni-
toring, implementing models, and ensuring the security of separations equipment
and processes. The preliminary development of the Digital Twin is taking place at
INL’s Solvent Extraction Laboratory, where two teams are concurrently engaged in its
development. One team is integrating conventional sensors, such as optical sensors,
while the other team is incorporating nontraditional sensors, such as acoustic sensors.
To achieve the ultimate objective of a fully functional Digital Twin, various compo-
nents are being integrated into a unified system through the creation of adapters.
These adapters connect a central data repository, DeepLynx, to various project
nodes, including LabView data streams, machine learning models, chemical models,
international safeguards support, and visualizations.
The fusion of Digital Engineering and Digital Twins (DTs) is revolutionizing the
nuclear sector, transforming product conceptualization and design and leading to
an era marked by increased efficiency, precision, and innovation. These initiatives
not only deepen our grasp of sensor integration and analytical methods but also
yield modular components that are adaptable for a variety of uses. Additionally, they
228 J. Darrington et al.

establish a template for data-driven evaluations that are instrumental in the creation
and oversight of emerging aqueous metal separations technologies.

2.2 Design

DTs offer transformative potential for nuclear reactor design, particularly during the
basic and final stages. These sophisticated virtual models can be instrumental in eval-
uating both the reactor’s nuclear components for safeguards and the Nuclear Power
Plant’s (NPP) overall security. A pivotal role of a DT at this stage is to ensure the
design incorporates the necessary infrastructure for security and safeguards moni-
toring, including strategic placement of sensors within the reactor core and facility
to maintain operational knowledge.

2.3 Security

Moreover, DTs can facilitate the early identification of security vulnerabilities,

enabling the International Atomic Energy Agency to address these issues proactively,
thus mitigating the expense of subsequent modifications. As a design progresses, a
specialized DT can be crafted to serve a dual role, bridging operational processes
with domestic and international safeguard and security protocols.
Operational DTs enhance the efficacy of both international and domestic inspec-
tions by pinpointing potential discrepancies for on-site verification and help maintain
regulatory compliance. They enable real-time monitoring, promptly notifying author-
ities of abnormal events. Crucially, DTs inform security strategies, differentiating
between deceptive tactics and actual threats, thus guiding responsive measures—
a consideration increasingly relevant for remote security management of small
autonomous reactors. The use of DTs allows access to information that is critical
for accurately predicting the diversion or misuse of a significant quantity of fuel,
through the collection of control rod data and a selection of assembly power over
the span of a year for sodium fast reactors [5]. For PBR reactors, sufficient statistical
analysis of diversion scenarios can be determined using gamma analysis [6].
In addition to advancements in nuclear safeguards, researchers have been
exploring the potential of advanced sensors and their ability to transmit data directly
to both a monitoring agency and the reactor itself [7]. Research in the field of indus-
trial control systems has the potential to enable a framework in which both the reactor
operator and the IAEA can receive independent streams of information. While this
research may not be directly related to safeguards, it plays a crucial role in elucidating
the intricacies of sensor-based monitoring for Small Modular Reactors (SMR) and
Medium Reactors (MR) systems.
Previous collaborations with the IAEA inform optimal sensor placements, bene-
fiting both operational safety and compliance monitoring. Thus, a DT embodies
The First Real-Time Digital Twin of a Nuclear Reactor 229

an integrated approach to the nuclear sector’s safety, security, and safeguards,

commonly referred to as the ‘3S by design’ concept, enhancing oversight and
ensuring continuous improvement.

3 Physical Description of the AGN-201

The AGN-201 is a compact nuclear reactor with a power output of 5.0 watts, situated
at Idaho State University. This reactor has been designed for both research purposes
and practical training [8]. Its structure comprises two primary components: the core
region and the ex-core elements.
The core region encompasses nine fuel disks that consist of a homogeneous
mixture of polyethylene and uranium dioxide fuel, enriched uranium. The core’s
approximate dimensions are 24 cm in height and 25.6 cm in diameter [68]. The core
is divided into two sections connected by a thermal fuse made of polystyrene, which
contains twice the fuel loading of the fuel disks. The thermal fuse serves as a safety
feature, it will melt and separate the core into two pieces, thereby terminating the
chain reaction if the temperature rises above a specified threshold.
To sustain criticality, four fuel control rods are inserted into the core. Two safety
rods remain fully always inserted, while a coarse control rod (CCR) and a fine control
rod (FCR) are employed to adjust, reduce, or maintain the reactor’s power output.
Additionally, there is a central irradiation facility passing through the core’s center,
allowing for experiments to be directly inserted into the core.
The core is enclosed by a graphite reflector contained within an aluminum core
tank. Surrounding this tank is an additional layer of graphite reflector, followed by a
lead shield, and finally, a water shield. Above the core, there is a removable graphite
thermal column that can be taken out for experimental purposes, and below the core
are the control rod drive mechanisms. The complete configuration of the AGN-201
reactor is illustrated in Fig. 1.

4 Risks and Hurdles, Challenges in Implementing Digital

Twins

There are numerous advantages of using DT technologies. They contain the ability to
encompass many of the necessary international safeguards and security of reactors,
but there are also challenges that arise when implementing a DT. Using any digital
product such as a Digital Twin, artificial intelligence, and machine learning, risk and
uncertainty are introduced into the system, and can possibly become a high risk.
Some of the challenges of using these products are:
1. Absence of regulatory guidance and requirements for compliance and acceptance
of use.
230 J. Darrington et al.

Fig. 1 AGN-201 reactor, reflector, and shielding core

2. Absence of any meaningful acceptance of international standards for design

(including cybersecurity) of DTs, AI, and ML.
3. Ensuring the workforce has the essential skills to design, evaluate, implement,
sustain, and update technologies as needed.
4. Performing the proper verification and validation of its performance over the
facility lifecycle.
5. Assuring user trust through proof, explanations and recurrent performance
testing/evaluation.

5 Technical Implementation

Digital Engineering and the development of digital twins necessitate the use of highly
sophisticated tools and software [26]. DTS allow the early detection, deterrence, and
response to potential threats by leveraging AI/ML algorithms for real-time moni-
toring of reactor data streams. Idaho State University and the Idaho National Lab
partnered up to create a digital twin centered around the AGN-201 reactor.
The First Real-Time Digital Twin of a Nuclear Reactor 231

The digital twin is composed of several components. First, data is acquired

through the Data Acquisition System (DAS) and uploaded into the open-source
INL-developed data lake DeepLynx1 through the INL-developed command line tool
called Jester. Once data has been ingested into the data lake, various machine learning
and AI operations occur on the data and feed their results back into the data lake.
These results are available on the data lake’s web interface or through a beta operator
windows program also developed by INL. Additionally, this data fuels user inter-
action through augmented reality and a real-time model of the reactor and the data
through INL-developed augmented reality programs.
Each section of this chapter will highlight a specific part of the diagram below, as
well as diving into the reasons behind certain technical decisions.

5.1 Jester

Jesteris an open-source tool developed by Idaho National Laboratory that transfers

data from sensor systems to the DeepLynx data lake. This software was developed and
open-sourced as part of the ISU digital twin efforts but was designed with modularity
and extendibility in mind. Jester has an extensive plugin system which allows it to
work on a myriad of different systems and with different file types.
Both Jester and an ISU specific plugin were developed as part of this effort.
This plugin is responsible for working with Jester to inform how the AGN-201’s
data acquisition system outputs the sensor readings and how to take those readings
and ingest them into DeepLynx. This plugin is written in Rust and is designed to tail
various CSV files output by LabVIEW and send them on an interval to the DeepLynx
data lake.
Rust
Rust is a multi-paradigm, general-purpose programming language that emphasizes
performance, type safety, and concurrency. It enforces memory safety, ensuring that
all references point to valid memory, without requiring the use of a garbage collector
or reference counting present in other memory safe languages. Rust’s rich type system
and ownership model guarantee memory safety and thread-safety—enabling you to
eliminate many classes of bugs at compile-time.
Rust was the preferred language due to having a smaller learning curve compared
to C++ , and a faster processing time than Python. Having a fast runtime and memory
safety are extremely important since the AGN-201 reactor’s DAS machine is small
and has limited memory. Choosing the Rust programming language allowed ISU
and INL the ability to safely and securely work with a nuclear reactor to power the
digital twin.

1 https://github.com/idaholab/Deep-Lynx.
232 J. Darrington et al.

5.2 DeepLynx

DeepLynx is an open-source data lake focused on enabling complex projects to

embrace digital engineering. Unlike other data warehouses, DeepLynx users can use
an ontology to custom-define how their data will be represented. DeepLynx enables
users to store their data in a graph-like format, ensuring that connections between
data can be easily seen and understood. This data lake also allows users to store
tabular, or time-series, data such as the data coming in off of sensors and IOT.
DeepLynx is hosted on Microsoft Azure for Government cloud platform lever-
aging the Azure Kubernetes system (AKS) to deploy and manage infrastructure. The
Microsoft Azure cloud platform grants DeepLynx the ability to connect all other
processes and software developed for this digital twin (Fig. 2). It is important to note
that while fast, this process is not considered “real-time” but “near real-time” due to
the network latency between each process communicating over the web (Fig. 3).
PostgreSQL
PostgreSQL a free and open-source object-relational database management system
(RDBMS) that can store and load data over time through associated pairs of times and
values. PostgreSQL possesses an incredible number of features to enhance perfor-
mance, security, programming extensions, and configuration. PostgreSQL powers

Fig. 2 Data flows into the data acquisition system from the reactor. Jester is used to push data from
the DAS into a DeepLynx cloud instance. The data is analyzed using machine learning algorithms
implemented through Juypter notebooks managed through Papermill, and pushes those results back
into DeepLynx, which then can be viewed by users through DeepLynx’s UI or the operator Window’s
program
The First Real-Time Digital Twin of a Nuclear Reactor 233

Fig. 3 DeepLynx representation of the AGN-201 in graph format. This representation of the data
allows a human readable view of the time-series data collected from the AGN-201 reactor

data mining to obtain patterns or models from the gathered data from all kinds of envi-
ronments [25]. It has the power that relational database management systems have
with the bonus of object-oriented features such as defining complex data types, over-
loading functions to work with different argument data types, and defining inheritance
relationships between tables.
Timeseries Data
DeepLynx is designed to not only store a graph representation of the AGN-201
reactor, but also serves as the storage for the data read from the sensors themselves,
as previously stated. Currently, DeepLynx leverages TimescaleDB for its storage
and querying of time-series or tabular data. This is another open-source technology
which was adopted by the DeepLynx team.
Data is processed only in near real-time due to network connectivity via HTTPS.
The different components of Fig. 1 are connected via TCP sockets using the HTTPS
format which is not necessarily the best for high throughput data. This stops the
system from producing real-time results. Currently the system runs in about thirty
second loops, including the machine learning and AI portions of the flow (Figs. 4
and 5).
234 J. Darrington et al.

Fig. 4 DeepLynx representation of data from the Multilayer perceptron algorithms

Fig. 5 DeepLynx representation of AGN-201 Engineering sensor data

5.3 Visualization Technology

We have made significant progress in realm of digital twin visualization for the AGN-
201 project. We have developed programs using various tools to enable operators
and monitors to easily identify anomalies and issues in reactor operation.
The First Real-Time Digital Twin of a Nuclear Reactor 235

Fig. 6 The operator UI allows operators to view trends and monitor the reactor in a human readable
format. This is a screenshot of the Windows desktop application

Operator UI
The operator UI is a beta desktop application designed to give insights into reactor
function to a reactor operator and monitor. This system incorporates the machine
learning data ingested by DeepLynx from the program listed previously and displays
it to the end user in an easy to digest format. This desktop program will allow
operators to digest the data from the prediction models and use that information to
understand risks and problems in the reactor proactively. The operator UI was created
by using a combination of Tauri,2 an open-source tool that builds optimized, secure,
and frontend-independent application for multi-platform deployment, and Svelte,3
an open-source front end component framework/language (Figs. 2 and 6).

2 https://tauri.app/.
3 https://svelte.dev/.
236 J. Darrington et al.

Augmented Reality Using Unity and HoloLens

Microsoft HoloLens is an augmented reality (AR)/mixed reality (MR) headset with
positional tracking technology developed and manufactured by Microsoft. In combi-
nation with Unity, a cross-platform game engine, an AR interactive view of the
AGN-201 reactor can be used by operators to see data that flows in. The AR systems
pulls in the predictions for the raw data from Deep Lynx and displays it through the
HoloLens using unity.

5.4 Machine Learning

A key aspect of the AGN-201 digital twin is the machine learning and artificial
intelligence detection programs. These programs take the live data from the reactor,
pulling it directly from DeepLynx, and performs various tasks to determine different
things. We will discuss the actual models and mathematics later on this chapter, for
now we will focus only on the technology involved in the operations.
Papermill
Papermill4 is an open-source tool that allows you to execute Jupyter Notebooks an
parameterize them. This allows the execution of workflows without manually having
to combine different notebooks. INL-developed a tool in Rust to manage Papermill’s
execution of various Jupyter notebooks (Fig. 1, Sect. 2.1.1). This allows the system
to use the memory efficiency/security of Rust, while leveraging the power/flexibility
of Python for machine learning. This allows us to maintain a highly available, highly
resilient system while at the same time letting data scientists work in a familiar system
and paradigm. This system is contained in a soon to be open-sourced Docker5 image.
DuckDB
DuckDB6 is an in-memory process/in process SQL OLAP database management
system. It leverages columnar storage which makes it ideal for analytics and allows a
simpler data format that data scientists can read. We utilize DuckDB as a temporary
store for the data from DeepLynx, putting it into an easily queried and read paradigm
that allows data scientists to focus on their algorithms instead of the data fetching
and network concerns of communicating with DeepLynx.

4 https://papermill.readthedocs.io/.
5 https://www.docker.com/.
6 https://duckdb.org/
The First Real-Time Digital Twin of a Nuclear Reactor 237

5.5 Conclusion

INL and ISU was able to build a DT with near real-time system for anomaly predic-
tion and general analysis of the AGN-201 reactor. There were several lessons learned
from this process that will continue to advance the understanding of how DTs are
implemented. First, the DT is near real-time due to the limitations of data trans-
mission through https and security firewalls. This constrained the DT to leverage
batch processing instead of streaming the data from the AGN-201 reactor. Second,
rust is a newer language and needed a lot of debugging to establish a dependable
interop between rust and python to effectively manage and analyze reactor data being
processed by Deep Lynx.

6 Physics Model Design “Physics-Based Surrogate Models”

Citation Key: [2] = ~ \cite{scikit-learn}.

The creation of a surrogate model (SM) for simulating a nuclear reactor involves
two main components: a mathematical model for reactor physics and a simplified
model for the reactor’s kinetic equations (PKE-SM). This model gets its information
from extensive simulations that mimic both normal and abnormal reactor operations.
Using a high-fidelity physics modeling software called Serpent, these simulations
consider factors like water temperature and control rod positions to determine the
reactor’s status—whether it’s operating normally (critical), shut down (subcritical),
or in an unsafe condition (supercritical). The results from these simulations are fed
into algorithms to create the SM, which can predict the reactor’s behavior based on
those inputs.
During the reactor’s operation, the SM continuously checks conditions like water
temperature and control rod positions to ensure the reactor is functioning as expected.
If these parameters stay within a safe range, it signifies that the reactor is in a stable,
critical state (Fig. 7).

6.1 Model Framework

The model framework consists of two primary components: (1) Model Development
and (2) Model Implementation and the models attempt to predict the following: Ch2
(Watts), Ch3 (Watts), FCR (cm), and CCR (cm).
Our methodology in machine learning (ML) serves as an automated system for
quality assurance and control (QAQC), as well as a framework for probing different
methods. The core of our strategy is the formulation of models that predict outcomes
for critical components such as Coarse Control Rods (CCR), Fine Control Rod
Heights (FCR), and the power outputs of Channel 2 and Channel 3 (measured in
238 J. Darrington et al.

Fig. 7 Flow of data from the AGN-201 through the mathematical SM and PKE-SM

Watts). We achieve this by employing all available variables, excluding the target
variable, to serve as predictors within our models. The predictor variables, detailed
in Table 1, include Channel 1 count rate (Ch1, CPS), power readings from Channel
2 (Ch2, W) and Channel 3 (Ch3, W), positional measurements of Fine Control
Rods (FCR, cm) and Coarse Control Rods (CCR, cm), Temperature (Temp, °C), and
Inverse Period (Inv-P, sec).
In this research, we have adopted three distinct model architectures, namely Multi-
layer Perceptron (MLP) Regression, GrowNet, and TabNet, each bringing a unique
perspective to the analysis of tabular data.
Multilayer Perceptron Regression, drawing upon classical neural network
paradigms, serves as a foundational model to set a benchmark for performance
comparison [9]. On the other hand, GrowNet is a novel gradient-boosted approach
that employs a composite of simpler neural networks, which are iteratively combined
to form an intricate and robust model adept at optimizing a variety of loss functions
[10]. A notable feature of GrowNet is its comprehensive corrective phase, which
fine-tunes the parameters across all contributing networks to enhance the model’s
predictive capacity.
TabNet, employing a sequential attention mechanism within a deep learning
context, stands out by sequentially parsing through variables to identify and prioritize
the most significant ones in relation to the overall dataset features [11]. This focus
on salient variables allows TabNet to make more informed predictions by effectively
allocating computational attention where it’s most impactful.
Together, these three methodologies represent the spectrum from tried-and-true
to state-of-the-art, encapsulating a broad range of capabilities in tabular data analysis
and providing a robust suite of tools for developing alongside digital twins.
The First Real-Time Digital Twin of a Nuclear Reactor 239

7 ML Model Development and Framework Design

The model development portion of the framework consists of the following steps:
1. Training Data Preprocessing: The training data set is cleaned using a rule-based
quality assurance/quality control (QAQC) along with a standard normalization.
First, the data is scrubbed by reviewing the data for errors that violate the reactor’s
known data thresholds, specifically identifying and fixing values that exceed
power maximum, and fall underneath reactor power minimums. This also consists
of populating any missing and incorrectly formatted missing data (NaN values).
Each variable is then adjusted using a standard normalization function. Finally,
the mean values and standard deviation values are stored for subsequent access.
2. Model Training: The training data is split into training, testing, and validation
datasets. This function is serialized so that a unique model is developed for each
variable.
3. Meta Data Storage: This step allows the model implementation step to access
these values for consistent data normalization.
4. Trained Model Export: Each trained model is exported. The twin can accom-
modate any file format that is appropriate for a specific model type.

7.1 Model Implementation

The near real-time model implementation portion of the framework consists of the
following steps:
1. Data Ingestion: The framework incrementally ingests data via SQL queries,
formats it to match the original training data, and stores the information in
memory.
2. Data Preprocessing: The framework isolates each data time step individually
and performs an automated QAQC and standard normalization using the mean
and standard deviation values it retrieves from the meta data produced during
model development.
3. Prediction: The trained model files are accessed and used to predict the values
of each variable (FCR, CCR, water temperature, Ch1 power, Ch2 power, Ch3
power) for the isolated and preprocessed data set.
4. Denormalization: The model predictions are returned to their original units by
undoing the normalization function.
5. Result Storage: The model results include the reported and predicted values
for each variable along with how those values differ. Under normal conditions
while the reactor is critical, the reported and predicted values should be nearly
identical. Deviation between those values is indicative that either (1) additional
model training is required, (2) sensor failure, or (3) the reactor is not operating
240 J. Darrington et al.

under expected conditions. The model does not attempt to identify which of those
three scenarios is taking place, but the discrepancy is stored in the results so that
it can be investigated further.

7.2 Computation and Using High Performance Computing

(HPC)

Creating a nuclear safeguards digital twin leverages the combined power of cloud
computing and high-performance computing (HPC). While the bulk of the digital
twin safeguards reside on the cloud, they collaborate with the INL HPC for opti-
mization algorithms and physics computations. This setup endows the digital twin
with the robust computational strength of HPCs and the flexibility and adaptability
of cloud storage.
Simultaneously, a groundbreaking digital twin has been developed, tested, and
validated for the autonomous control of live assets within the Deep Lynx ecosystem.
This advancement, funded through the INL Laboratory Directed Research and Devel-
opment (LDRD) initiative, incorporated tools like Deep Lynx, the ML adapter,
and physics simulations grounded in the Multiphysics Object-Oriented Simulation
Environment (MOOSE). The INL’s Microreactor Agile Non-Nuclear Experimental
Testbed (MAGNET) serves as a testing platform for microreactor technologies.
Notably, on March 30, 2022, the LDRD digital twin was employed in a MAGNET
test involving a single heat pipe. The test showcased the digital twin’s prowess in
forecasting temperature benchmarks and adjusting MAGNET’s temperature set point
autonomously, enabled by the seamless two-way communication between Deep Lynx
and the MAGNET data acquisition system.
The digital twin is slated for hosting on a blend of the Microsoft Azure cloud and
HPC infrastructure. Given the rigorous computational requirements of the Serpent
simulations, an HPC environment is best suited for rapid simulation processing.
While the Serpent simulation is contained within the INL HPC cluster, other aspects
of the digital twin will be housed in the Azure cloud. This arrangement offers
several advantages, such as scalability, cost-efficiency, enhanced security, and deeper
insights into hardware and network dynamics. For seamless data exchange between
the cloud and HPC, specific components and a data-flow framework are imperative.
For instance, Deep Lynx in the cloud can initiate an HPC processing request. An
HPC adapter fetches required files from Deep Lynx and secures them in the Azure
storage. Within the HPC, a service node constantly scans the Azure storage for fresh
requests and related files. Post-processing, results get transferred back to the cloud
storage. This HPC adapter consistently monitors the storage, relaying results back
to Deep Lynx once detected.
The First Real-Time Digital Twin of a Nuclear Reactor 241

8 Visualization

Extended Reality (XR) encompasses a suite of technologies including Virtual Reality

(VR), Augmented Reality (AR), and Mixed Reality (MR) that merge the virtual and
physical worlds. VR immerses users in a completely virtual environment, whereas
AR enhances reality with virtual overlays, and MR combines the two to facilitate intu-
itive interactions between human, computer, and environmental elements (Milgram
and Kishino, 1994). Each of these technologies offers distinct user experiences. VR
immerses users in a completely virtual environment, disconnecting them from the
physical world. Idaho National Laboratory (INL) leverages these XR capabilities in
projects that span Nuclear, Integrated Energy, and National and Homeland Security
missions (Idaho National Laboratory, n.d.).
At the forefront of Digital Engineering, INL integrates visualized data streams,
visualizations, and physics-based models to create comprehensive digital twins.
These digital twins enable semi-autonomous design, autonomous operation, and
real-time anomaly detection, thus transforming complex data into accessible and
actionable insights.

8.1 The Advantages of Incorporating XR for Data Analysis

Extended Reality, often abbreviated as XR, is an umbrella term encompassing various

immersive technologies, such as Virtual Reality (VR), Augmented Reality (AR), and
Mixed Reality (MR).
The primary advantage of integrating XR into data analysis processes lies in its
ability to facilitate the interpretation and representation of complex data. Traditional
data analysis techniques often rely on subject matter experts to manually collect,
examine, and report anomalous data. By providing operators with a spatial view
of captured data, XR simplifies the process of identifying and plotting data points.
This enhancement makes data and anomalies more engaging, enabling operators to
readily identify unknown signals in real-time or near real-time.

8.2 QR Code Scanning for Precision

QR code scanning plays a significant role in our workflow by enabling precise posi-
tioning of virtual assets relative to the physical world. This process allows engineers
to conduct design reviews, test ergonomics, fitment, escape routes, and address prac-
tical challenges, ultimately enhancing the design and testing phases (Idaho National
Laboratory, 2023).
242 J. Darrington et al.

Our interaction libraries, integrated via SDKs, enable us to manipulate objects,

create interactive menus, toggle layers, and utilize QR code scanning for precise
model positioning within physical spaces. This spatial modeling approach facilitates
design reviews, ergonomic assessments, and fitment evaluations. It also addresses
unforeseen challenges, such as navigating tight spaces, thereby enhancing the overall
design process. This integration allows the visualization of dynamic states such as
flow rates, temperature heat maps, and fuel rod positions, thus eliminating the need
for intermediate data interpretation (Smith et al., 2021).

8.3 Unity and OpenXR

Utilizing the game creation engine Unity (Unity Technologies n.d.), INL has devel-
oped a robust environment for visualizing digital twins. This platform was chosen
for its versatility in building projects across various platforms and its supportive
community.
INL begins by converting Computer-Aided Design (CAD) models into a mesh
that is easier to render using PIXYZ, a tool designed to optimize models by simpli-
fying geometry without compromising on detail (Unity Technologies n.d.-a). The
primary advantage of integrating XR with data analysis lies in its capacity to facilitate
the interpretation and representation of intricate datasets. Conventional approaches
often rely on subject matter experts to manually collect, examine, and report anoma-
lous data. Our approach, however, offers spatial visualization capabilities, simpli-
fying the process of data identification and plotting. This real-time or near real-time
engagement enables operators to swiftly detect and understand unknown signals and
anomalies, elevating our decision-making capabilities.
While HoloLens 2 remains our primary choice, the industry’s direction empha-
sizes cross-platform compatibility. OpenXR platform enables headset makers to
define hardware specifications while utilizing standardized interaction methods,
ensuring that applications can be developed for multiple headsets with ease (Idaho
National Laboratory 2023). Following the conversion process, these assets are
imported into Unity, where INL engineers craft interactive experiences tailored for
XR headsets like the HoloLens 2 (Microsoft n.d.), known for its Mixed Reality Toolkit
(MRTK) that facilitates the development of immersive applications with features
such as hand tracking, hand poses, speech recognition, eye tracking, gaze tracking,
speech recognition, and other interaction methods (Microsoft n.d.-a). Alternatively,
the industry is moving toward the OpenXR standard, which promises cross-platform
compatibility for XR applications (The Khronos Group n.d.).
The First Real-Time Digital Twin of a Nuclear Reactor 243

8.4 Connecting Real-Time Data with Digital Twins

To achieve a comprehensive digital twin, we interconnect real-time data from existing

assets, a process successfully applied to MAGNET and the ISU Twin. Utilizing
data acquisition (DAQ) systems and importing data into DeepLynx is our preferred
method. DeepLynx seamlessly integrates data, providing real-time insights into
system states. XR headsets eliminate the need for data interpretation, allowing
users to visualize the system’s status intuitively. This visualization includes flow
rates, temperature or concentration heat maps, fuel rod positions, conflict areas, and
dynamic animations of construction processes. Thanks to DeepLynx, these visual-
izations can be dynamically updated as new data flows in, accessible across various
platforms.
Unity’s “GameObjects” and their associated components and scripts, written in
C#, facilitate the creation of interactive digital twins. These objects can represent
various 3D objects, text, or abstract constructs within the scene hierarchy (Unity
Technologies n.d.-b).

8.5 Collaborative XR Experience with Serval Multiplayer

An innovative feature utilized by INL is the Serval Multiplayer server, which enables
collaborative sessions by synchronizing state data among multiple users, even though
they operate in individual sessions. By allowing users to scan the same QR code, they
can view models in synchronized positions while maintaining individual sessions,
ensuring privacy. This enhances the collaborative process, allowing simultaneous
interaction with and visualization of the digital twin (Serval Project n.d.).
In conclusion, INL’s integration of 3D visualization and XR technologies into
digital twins represents a significant advancement in the field, offering enhanced
data interpretation, real-time operational insights, and improved collaborative
capabilities.

9 Case Study of the Impact the AGN-201 Digital Twin: Red

Versus Blue Test and Conclusion

Necessary Terminology
CCR Coarse control rod.
DT Digital twin.
FCR Fine control rod.
IFML Isolation forest machine learning.
INL Idaho National Laboratory.
244 J. Darrington et al.

ISU Idaho State University.

ML Machine learning.
PKE-SM Point kinetics equations surrogate model.

9.1 Overview

In the past year, the reactor operators at Idaho State University (ISU) operated the
AGN-201 reactor under various conditions without informing the digital twin (DT)
team at Idaho National Laboratory (INL) to perform both offline and online detec-
tion of hypothetical proliferation attempts. These data were transmitted to DeepLynx
during the operation and stored for analysis by the DT team. During this time, knowl-
edge of the actual operations was kept from the DT team to allow for an offline anal-
ysis and estimate of events. Three separate analyses (reactor physics assessment,
machine learning predictions, and anomaly detection) were performed by the DT
team to examine the DT’s ability to detect and quantify off-normal operations. This
provided first-of-a-kind online and offline monitoring of a nuclear reactor using a
DT, proves the impact, and necessity of the utilization of digital twin’s in the realm
of security, and validation of critical systems.

9.2 AGN-201 Operations

The red/blue team test involved the ISU reactor supervisor (red team) modifying
the operational parameters of the AGN-201 reactor, within regulatory limits but in
ways the DT had not previously seen, to simulate a potential bad actor nefariously
gaining access to the reactor to produce a dangerous isotope or other illicit use. The
DT analysts (blue team) then analyzed the resulting outputs from the reactor and
determined if/when anomalous actions were taken.
The test consisted of three separate alterations to the reactor with different amounts
of reactivity change. From highest to lowest reactivity change, the list of alterations
to the reactor is:
. Inserting a polyethylene rod into the central irradiation facility.
. Inserting a cadmium foil cover into the central irradiation facility.
. Removing two graphite blocks from the thermal column.
All the alterations were performed while the reactor was operating, and the
operator and reactor supervisor tried to make these changes as hidden as possible.
After a sequence of first determining if the reactor was critical using a fine control
rod (FCR) and a coarse control rod (CCR), a “nefarious” was emulated by placing
cadmium foil cover into the central irradiation facility to act as a poison. The cadmium
was inserted into the reactor approximately 20 min after the polyethylene rod was
removed, resulting in the FCR.
The First Real-Time Digital Twin of a Nuclear Reactor 245

9.3 Online Detection of Reactor Operations

To enhance the Digital Twin (DT) of the AGN-201 reactor with an effective anomaly
detection system, the team implemented the Isolation Forest machine learning algo-
rithm during the Red versus Blue Test. This algorithm is particularly good at spotting
unusual patterns in data without the need for previously identified examples of such
anomalies.
The team first standardized the data, then fine-tuned the Isolation Forest model to
the specific characteristics of the DT’s data. They also set a ‘contamination parameter’
to estimate the expected ratio of anomalies in the data.
As the reactor operates, the algorithm continuously analyzes incoming data,
deciding if each piece is normal or not. If it detects what seems to be an anomaly,
it marks and logs the occurrence for later review. This system is key in promptly
identifying and responding to any irregularities that might arise.
The initial operation was used to assess the acceptability of the reactor physics
model and anomaly detection model for predicting off-normal conditions (i.e., simu-
lated proliferation attempts). Given the success of both models, two additional reactor
operations were performed to assess the online capabilities of the DT system. This
ensured that data was being transferred from the AGN-201 digital acquisition system
to DeepLynx, where it could be analyzed in near real-time. This process allowed the
reactor operators at ISU to operate for approximately an hour, and then within the
next 20 min, the data could be assessed and discussed between the two teams. The
ability to provide a near real-time assessment of reactor operations was a first for a
DT in nuclear reactor operations.

9.4 Takeaways from Red Versus Blue Test

Overall, the initial red/blue team exercise was a success in showing that a Digital
Twin framework could effectively transfer data from AGN-201 via DeepLynx and
analyze it for off-normal operations. This was demonstrated for both offline and
online operations to provide confidence that the methodologies present would be
transferable to the near real-time detection of anomalies.
Through the creation of a digital twin, the reactor physics surrogate and anomaly
detection models were able to capture the major anomalous behaviors that the reactor
operations team performed. The integrated anomaly detection algorithm within the
twin displayed a remarkable aptitude in capturing many of the off-normal opera-
tions, homing in on sensor fluctuations to identify divergences from the norm. This
achievement is particularly noteworthy given the complexity of the reactor’s oper-
ational dynamics and the nuanced variations that were successfully flagged, show-
casing the algorithm’s precise tuning and responsive nature to dynamic operational
changes, and the impact of using DT’s in this realm.
246 J. Darrington et al.

Future work will entail wrapping in the PKE-SM to enhance the capabilities of
the reactor physics surrogate model. This will allow us to capture changes in power
as a function of time rather than just examining keff . For the reactor physics models,
an additional area of focus will be on determining what levels of perturbation are
quantifiable, meaning how small of a perturbation can be detected before it falls
within the range of uncertainty in the model, sensor data, etc. For both the reactor
physics surrogate and anomaly detection models, we plan to focus on detecting
startups, power changes, and shutdowns to prevent flagging these as anomalies.
Through further testing, we were able to provide a near real-time assessment of
reactor operations and were able to detect nearly all the anomalous behaviors of the
reactor.

9.5 Conclusion

The Digital Twin of the Idaho State University’s AGN-201 was a resounding success,
despite the difficulties and hurdles in both operation in implementation. The Idaho
National Laboratory team was able to successfully create a near real-time twin of a
functioning nuclear reactor, a first in the industry. Not only did they create a way for
monitoring the twin using readily available cloud services but worked toward creating
a framework for monitoring the correct usage of said reactor to stop potential bad
actors.
This work will lay the foundation for future digital twins in the Idaho National
Laboratory’s sphere of influence and has proven that they have the capability to
handle larger projects.

References

1. Ritter C, Hays R, Browning J, Stewart R, Bays S, Reyes G, Schanfein M, Pluth A, Sabharwall

P, Kunz R, Shields A, Koudelka J, Zohner P (2022) Digital twin to detect nuclear proliferation:
a case study. J Energy Resour Technol 144(10):03, 102108. [Online]. Available: https://doi.
org/10.1115/1.4053979
2. “Digital engineering strategy,” U.S Department of Defense, Tech. Rep. [Online]. Available:
https://man.fas.org/eprint/digeng-2018.pdf
3. Yadav V et al (2021) The state of technology of application of digital twins. U.S. Nuclear
Regulatory Commission, Tech. Rep. TLR/RES-DE-REB-2021-01, 2021.[Online]. Available:
https://www.nrc.gov/docs/ML2116/ML21160A074.pdf
4. Yadav V et al (2021) Technical challenges and gaps in digital-twin-enabling technologies for
nuclear reactor applications. U.S. Nuclear Regulatory Commission Tech Rep TRL/RES-DE-
REB-2021-17. [Online]. Available: https://www.nrc.gov/docs/ML2136/ML21361A261.pdf
5. Stewart R, Shields A, Gleicher F, Bays S, Schanfein M, Reyes G, Ritter C (2023) Utilizing a
digital twin to analyze and monitor proliferation scenarios in a sodium-cooled fast reactor. J
Sci Global Secur (Submitted)
The First Real-Time Digital Twin of a Nuclear Reactor 247

6. Wen S, Stewart R, Shields F, Gleicher A, Bays S, Reyes G, Schanfein M, Ritter C (2023)

Utilizing advanced statistics to determine anomalistic conditions in pebble-bed reactors. J
Nucl Mater Manage (Submitted)
7. Browning J, Slaughter A, Kunz R, Hansel J, Rolston B, Wilsdon K, Pluth A, McCardell D
(2022) Foundations for a fission battery digital twin. Nucl Technol 208(7):1089–1101. [Online].
Available: https://doi.org/10.1080/00295450.2021.2011574
8. Pope C, Pheonix W (2021) Idaho State University AGN-201 low power teaching reactor—
An overlooked gem, ser. nuclear reactors spacecraft propulsion, research reactors, and reactor
analysis topics
9. Safety analysis report of the idaho state university agn-201m research reactor (2003). Tech.
Rep. License No. R-110, Docket No. 20–284
10. Sercan A, Pfister T (2021) TabNet: attentive interpretable tabular learning. Proc AAAI Conf
Artif Intell 35(8):6679–6687. https://doi.org/10.1609/aaai.v35i8.16826
11. Murtagh F (1991) Multilayer perceptrons for classification and regression. Neurocomputing
2(5–6):183–197
Evolution of Simulation and Digital Twin
in Health Care: From Discovery
to Design and Integration

Yue Dong , Amos Lal , Alexander S. Niven , and Xiang Zhong

Abstract Over the past two decades, simulation and Digital Twin (DT) technolo-
gies have become increasingly prevalent in health care. These technologies have
significant potential to advance modern medicine, enhance clinical decision-making
and team performance, improve healthcare delivery, reduce cost, and improve
patient outcomes. This chapter provides an overview of these technologies and
their emerging applications in the field of health care, including opportunities to
accelerate discovery in basic science, delivery of more realistic training opportu-
nities that advance clinician competence and interprofessional teamwork, and an
efficient and cost-effective approach to analyze, improve, and monitor clinical work-
flows, healthcare delivery systems, and their performance. This emerging field fosters
multidisciplinary research among healthcare professionals, information technology
experts, engineers, and data scientists, all working together to better serve our society.
Widespread adoption of these technologies will require solutions to address several
technical, ethical, and regulatory challenges. These solutions will require a close
collaboration between industry, academic centers, and government to develop a
thoughtful approach that aligns implementation and integration of simulation and
DT technologies into the healthcare workplace with measurement of meaningful
outcomes to ensure broad-based access benefit from these important tools.

Y. Dong (B) · A. Lal · A. S. Niven

Mayo Clinic, 200 First Street SW, Rochester, MN 55905, USA
e-mail: [email protected]
A. Lal
e-mail: [email protected]
A. S. Niven
e-mail: [email protected]
X. Zhong
Univesity of Florida, Gainesville, FL 32603, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 249
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_11
250 Y. Dong et al.

1 Introduction

Modern health care is a complex adaptive system that requires highly skilled inter-
professional team members to work together using advanced equipment and standard
processes to care for patients with a wide variety of diseases with steadily increasing
acuity over time. While the rapid pace of advancement in medical science offers
exciting possibilities and new hope to patients and their family within our global
community, it also demands constant changes in an already complicated and strained
healthcare delivery system. The practice of medicine and current healthcare outcomes
fall far below the level of quality and safety achieved by high-reliability industries,
such as aerospace and manufacturing. As a result, institutions have become increas-
ingly focused on the “value equation” of the care they deliver, generally defined as
the quality of care divided by the total cost of patient care over time [1].
Over the past two decades, simulation and Digital Twin (DT) technology and
their applications have become increasingly prevalent in health care. These tools
can provide increasingly accurate representations of the pathobiology behind patient
conditions, opportunities to monitor patient health and optimize disease diagnosis and
management, enhance and evaluate the performance of clinicians and their healthcare
teams, and analyze and improve healthcare delivery and systems. Using a systematic
implementation strategy that includes key stakeholder collaboration, these technolo-
gies offer the opportunity to accelerate medical research discovery and improve
healthcare system performance and patient outcomes with greater efficiency and at
a reduced cost. This chapter provides an overview of these technologies and their
emerging applications in health care.

1.1 Healthcare Evolution from Individual Treatment

to Health Delivery Integration: A Systems Perspective

The twentieth century witnessed a wide range of groundbreaking scientific discov-

eries in medicine, including antibiotics, anesthesia, blood transfusions, vaccines,
Computed Tomography (CT) scans, magnetic resonance imaging (MRI), and extra-
corporeal therapies, including dialysis and extracorporeal Membrane Oxygenation
(ECMO), and more [2]. The twenty-first century promises an even faster pace
of discovery, with rapid advances in genomics offering a customized approach to
common and complex diseases using precision medicine, and the digital revolution
challenging our traditional approach to care delivery.
These advancements have had a profound impact on human health and have
significantly shaped both the current practice of modern medicine and the bene-
fits it provides to society. Healthcare delivery systems have evolved into complex
ecosystems of clinical practice environments that offer a wide variety of services
Evolution of Simulation and Digital Twin in Health Care: From … 251

to scale. However, these healthcare improvements come with their own set of chal-
lenges. Population aging and growth, the increasing prevalence of chronic condi-
tions, and advanced diagnostic and treatment modalities are both driving cost and
reducing access to and equity of care. As the complexity of disease management
has increased, there is a growing demand for a collaborative, multidisciplinary, and
team-based approach that includes physicians, nurses, pharmacists, and respiratory
therapists and technicians. These teams require a clear understanding of each other’s
roles and responsibilities and a high level of communication and teamwork to ensure
timely and coordinated care delivery.
Given its complexity, health care today has evolved into a “system of systems” of
healthcare delivery focusing on not only just safe and timely medical care, but also
better management of resources to meet demand at scale across geographic locations
and time. The Systems Engineering Initiative for Patient Safety (SEIPS) 3.0 frame-
work could serve as a guiding model in this complex landscape [3]. It offers a holistic
approach to integrating various components such as people, tasks, tools, and orga-
nization, thereby providing a comprehensive view of healthcare delivery systems.
These “systems” are also often stratified using different levels, including biology and
disease, patients and clinicians, and care delivery processes at the bedside, within
the hospital, and throughout healthcare systems as a whole [4]. A System of Systems
(SoS) in health care refers to integrating and interacting various independent systems
within the healthcare sector to achieve a more comprehensive, efficient, and effective
healthcare delivery.
In complex healthcare systems, the overall system performance is directly linked
to the functionality of each subsystem and the interconnections between them. Indi-
vidual components, such as patient care, administrative processes, and healthcare
technology, must not only function effectively in isolation but also integrate seam-
lessly. The synergy between these subsystems is crucial for the overall system perfor-
mance. Medical errors and gaps in care quality have been well documented. However,
fixing them is not easy due to the complex and unpredictable work environment
common in medical practice [5, 6]. The recent global COVID-19 pandemic and its
impact further exposed these persistent vulnerabilities in healthcare systems world-
wide [7–9]. The pandemic has served as a catalyst for change, underscoring the
importance of improving current healthcare delivery mechanisms. It is imperative to
enhance the efficiency of healthcare delivery while maintaining the highest standards
of quality care and use innovative solutions and technologies to improve healthcare
access and distribution on a global scale. This systems view of healthcare delivery
also provides a framework that we can use to better understand how simulation and
DTs can be used to offer important solutions to these complex problems.

1.2 Technology Advancement in Health Care

Over the past 30 years, we have witnessed a technological revolution that has signif-
icantly impacted all aspects of our lives, including health care. Personal computers,
252 Y. Dong et al.

mobile devices, the internet, cloud computing, and artificial intelligence/machine

learning (AI/ML) have led to an explosion of data related to biology discovery, disease
diagnosis, management, and healthcare delivery systems optimization. These data-
driven decision support tools can amplify and augment human capacity to improve
the efficiency, quality, and safety of health care during the whole journey of patient
experience (prevention, management, recovery, and wellness).
Simulation is the imitation or representation of one act or system by another. The
Society for Simulation in Healthcare describes four main purposes of simulation—
education, assessment, research, and health system integration to facilitate patient
safety [10]. Simulation-based training started in the discipline of anesthesiology and
has grown rapidly in other areas of medical education, with the development of
a growing variety of computer-controlled manikins and task trainers. These simu-
lators can be programmed to recreate a variety of standardized clinical scenarios,
allowing students to practice their clinical reasoning, technical (intubation, central
line insertion, etc.), and teamwork skills (communication) in a safe and controlled
environment. Simulation-based education has been shown to be an effective way to
improve student knowledge and skills and translate to better patient outcomes [11].
Computer simulation involves creating a computer-based model or virtual
representation of a system or process to study and analyze its behavior.
As an extension to traditional computer simulation, DT is a virtual representa-
tion of real-world entities and processes synchronized at a specified frequency and
fidelity. DTs’ systems transform business by accelerating holistic understanding,
optimal decision-making, and effective action. DTs use real-time and historical data
to represent the past and present and simulate predicted futures [12]. It is a digital
replica of a physical system that can be used to monitor, predict, and optimize the
system’s behavior. DT is a virtual representation of a physical asset replicated virtu-
ally through a data connection, making it possible to link the system with its virtual
counterparts in a bi-directional way. The “bi-directional” exchange of information,
which synchronizes the virtual system response to match the physical system, distin-
guishes DT from traditional digital simulation, which is often considered indepen-
dently operated. Specifically, a digital model has no interaction with the physical
system. A digital shadow is updated with data from the physical system but does
not inform its control. A DT arises when the physical system data is used to update
the digital model, and the resulting simulation is used to control the physical system
[13]. Very few works in literature have realized the control loop and are considered
as true DTs in use, and thus, our review also includes simulations that are digital
models or digital shadows [14].

1.3 Healthcare Applications

DTs have various applications in health care, such as training, diagnosis, manage-
ment, and care delivery for patients, providers, and healthcare organizations. One
Evolution of Simulation and Digital Twin in Health Care: From … 253

way to categorize DTs is based on Siemens’ framework, which has three levels of
twinning [15].
First, product twining provides a virtual-physical connection to analyze how a
product performs under various conditions and adjust in the virtual world to ensure
that the physical product will perform exactly as planned in the field. For product twin-
ning, one example is the DTs of medical devices (e.g., digital radiological devices)
[16, 17]. These efforts focus on the physical device/product twinning (similar to the
concept of DTs for management in manufacturing) and are mostly used to monitor its
status, diagnose issues, and test solutions remotely, ultimately optimizing its perfor-
mance and reducing the risk of malfunctions. From digital devices to digital/virtual
patients, precision medicine is one area that exemplifies the application of both
product twinning and process twinning. Digitally replicating the human body (from
cell to organ biological/physiological systems) allows for in silico clinical trials to
examine the prevention, early detection, and targeted treatments of many diseases.
These DT human body/organ systems have been used for drug development and
treatment recommendations [18]. Another example of virtual patients is optimizing
health care at the individual level, i.e., “personalized health monitoring,” with the
goal of healthcare management and promoting healthy behavior. It can be used to
create personalized models of patients that can be used to monitor their health status,
predict the course of disease, and optimize treatment plans [19]. This application
is still focused on modeling a virtual patient instead of the healthcare system at
large and is mainly focused on healthy people’s daily life and their living environ-
ment instead of a specific healthcare facility. The third level is system twinning,
which includes using DTs to improve hospital operation processes and workflows
by allowing managers to tweak inputs and see how outputs are affected without the
risk of upending existing workflow [20]. The system performance can be captured,
analyzed, and acted on operational data, providing insights for informed decisions
to maintain effective interactions among the components of the system at the system
level.

1.4 Outline of the Chapter

DTs have risen at the intersection of Industry 4.0 and the Internet of Things (IoT),
relying on converging technologies in big data analytics, pervasive sensing, and cloud
computing infrastructure. In this new era, simulation and DTs promise to enhance
healthcare delivery by improving efficiency, flexibility, and patient-centered care, all
of which are crucial in a post-pandemic world. This book chapter highlights how
technological advancements have enriched health care with valuable data and digital
assets, making it imperative to utilize these resources to optimize the entire spectrum
of healthcare services. We will delve into the applications, benefits, challenges, and
future prospects of DTs and simulation technologies in health care, particularly in
the context of a post-COVID-19 world acute care setting. We will explore how these
technologies can serve as invaluable tools for healthcare professionals, clinicians,
254 Y. Dong et al.

researchers, IT professionals, engineers, and data scientists, all working together to

serve society. In Sect. 2, we will explore simulation and DT applications in biology
and disease management. In Sect. 3, we will discuss the latest professional training
applications using simulation and DTs’ technology. In Sect. 4, we will present appli-
cations that support decision-making for hospital operations and management. Then,
in Sects. 5 and 6, we will provide a vision of simulation and DTs in health care and
discuss the challenges and opportunities.

2 Simulation and Digital Twin Applications at the Biology

and Disease Management

Modeling and simulation have emerged as a powerful tool for understanding complex
biological systems and processes in addition to the traditional theory and experiment
[21]. Utilizing mathematical models, simulation allows researchers and clinicians
to predict disease outcomes, test treatment interventions, and optimize healthcare
delivery. There are two common approaches for physiology and disease modeling:
rule-based and data-driven modeling. Each of these approaches offers distinct advan-
tages and challenges [22]. The rule-based approach relies on predefined rules and
knowledge about biological systems to model disease processes. These rules are
typically derived from clinical guidelines, expert opinions, or well-established phys-
iological pathways and experienced-based learning in clinical settings. It has the
benefits of transparency (by incorporation of directed acyclic graphs in the creation
and execution of rules), easy control of model behavior, and stability for output,
which make it easy to accept for clinicians [23, 24]. At the same time, it also has
stability and flexibility limitations because it requires expert knowledge and comes
with a possibility of bias. The mechanistic model is designed to represent the under-
lying mechanisms of a system—the processes and interactions that lead to observed
phenomena. They are often contrasted with empirical models, which are purely based
on observed data without assumptions about the underlying physiology processes.
The data-driven model uses large datasets to investigate inter-relationships using
different statistical and data mining methods. It is based on systems’ biology and
systems’ immunology and has expanded rapidly in the last decade because of more
accessible data from EMR and more computer power [25, 26].

2.1 Understanding Physiology and Treatment

The simulation technologies have revolutionized medical science by understanding

physiology. Various virtual training platforms now offer immersive experiences that
Evolution of Simulation and Digital Twin in Health Care: From … 255

closely mimic real-world clinical scenarios based on physiology [11]. These plat-
forms are designed to be interactive, allowing medical students and healthcare profes-
sionals to engage in problem-solving, diagnosis, and treatment planning in a risk-free
environment. The virtual training modules often include a wide range of case studies
covering various diseases, conditions, and patient demographics, providing a compre-
hensive learning experience. The lab and vital signs will change based on the inter-
vention by user input. The virtual patient offers a practical understanding of disease
mechanisms and treatment protocols, thereby bridging the gap between academic
learning and clinical practice. This hands-on approach enhances the learner’s ability
to make informed clinical decisions in a safe environment that will not harm patients,
ultimately leading to better patient outcomes.
Archimedes Diabetes Model was designed to simulate individual patients and
how they would respond to various treatments for diabetes. It incorporated a wide
range of variables, from glucose metabolism to treatment protocols, to generate
realistic patient outcomes [27, 28]. This model serves as one of the early successes
in utilizing computer-based physiology to understand and manage diabetes. This was
also an early implementation of using directed acyclic graphs to explain the complex
pathophysiology and rules based on the clinical environment. The model underwent
rigorous validation, and its predictions were found to be highly consistent with the
results of randomized controlled trials. This level of accuracy made the Archimedes
Diabetes Model a valuable tool for clinicians, researchers, and policymakers alike. It
not only helped in understanding the complex physiology of diabetes but also played
a crucial role in optimizing treatment plans and healthcare policies related to diabetes
management.
Agent-based models can simulate the interactions between various biological
agents, such as cells and molecules, in each biological system [29, 30]. These agent-
based models are particularly useful for understanding complex, multiscale biolog-
ical systems where traditional modeling approaches may fall short. By providing a
more nuanced understanding of biological systems, agent-based models have paved
the way for more targeted and effective treatments, bridging the gap between basic
biological research and clinical applications. This approach is offering insights that
are directly applicable to patient care and treatment optimization. The research group
led by Vodovotz explores the use of omics data and mathematical modeling to under-
stand and explore various critical illness management strategies by integrating data-
driven and knowledge-based modeling approaches [22]. This approach aims to inte-
grate large-scale data with computational modeling to improve research and clinical
applications in diseases involving severe inflammation and immune responses.
BioGears is an open-source, comprehensive human physiology engine that has
been instrumental in driving medical education, research, and training technolo-
gies [31]. BioGears aims to provide accurate and consistent physiology simulation
across the medical community. It can be used as a standalone application or inte-
grated with other simulators and sensor interfaces making it a valuable asset in the
healthcare simulation community. Similar to BioGears, the Pulse Physiology Engine
offers robust physics-based circuit and transport solvers [32]. It includes a common
data model for standard models and data definitions, a software interface for engine
256 Y. Dong et al.

control, and a verification and validation suite. Pulse’s architecture is designed to

reduce model development time and increase usability, making it a go-to solution
for many in the healthcare simulation community.
Emerging roles of virtual patients in the age of AI have been explored, and rule-
based systems have shown promise in healthcare modeling. These systems can simu-
late patient responses based on a set of predefined expert rules, making them useful for
training and diagnostic purposes. Dr. Lal and team have developed and verified a DTs
model of critically ill patients using the causal AI approach to predict the response to
specific treatment during the first 24 h of sepsis [23]. This expert rule-based DTs focus
on creating individualized patient models for better medical education and clinical
decision support [33]. This work was previously developed specifically for sepsis
in critically ill patients but has now evolved into multiple commonly seen clinical
scenarios in the intensive care setting. The primary goal is to provide a safe testing
bed for the learners and the clinicians to evaluate a proposed intervention (useful or
otherwise) in an in silico environment, before actually performing those interven-
tions on the patient at the bedside. The envisioned result is to avoid any preventable
harm to the patient by pre-testing the interventions that carry any uncertain risks. The
utility of such DTs in health care can serve multiple purposes, including advances in
medical education delivery, in silico research, and adjunct clinical decision support
at the bedside.

2.2 In Silico Clinical Trials

In silico clinical trials have emerged as a groundbreaking approach in medical

research, offering many opportunities and advantages over traditional clinical trials.
Utilizing computational models and simulations, in silico trials allow for the testing
of medical interventions or devices in a virtual environment, thereby revolutionizing
the way clinical research is conducted [34, 35].
One of the most compelling advantages of in silico trials is the significant reduction
in costs. Traditional clinical trials often require extensive resources, including patient
recruitment, site management, and long-term follow-ups, all of which contribute to
high operational costs. In contrast, in silico trials can be conducted with minimal
overhead, as they rely on computational power rather than physical infrastructure
and human resources. This cost-effectiveness makes it feasible to explore a broader
range of research questions and to conduct multiple trials simultaneously, thereby
accelerating the pace of medical innovation. Time is another critical factor that holds
a distinct advantage for in silico trials. Traditional trials can take several years to
complete, from initial planning to final data analysis. In silico trials, however, can be
executed in a fraction of the time. The ability to quickly simulate different scenarios
and interventions enables researchers to arrive at conclusions more rapidly, thereby
speeding up the “time to market” for new treatments and medical devices. Perhaps,
the most significant benefit of in silico trials is the elimination of risks to human
subjects. Despite rigorous ethical standards, traditional clinical trials always carry
Evolution of Simulation and Digital Twin in Health Care: From … 257

some risk to the participants. In silico trials remove this concern entirely, as they
are conducted in a virtual environment. This not only ensures the safety of poten-
tial patients but also allows for the testing of treatments that might be considered
too risky for traditional trials. By offering a safer, faster, and more cost-effective
alternative to traditional methods, in silico clinical trials are poised to become a
cornerstone in the future of medical research. They offer a promising pathway for
the development of new treatments and medical technologies, with the potential to
significantly improve patient outcomes and healthcare systems globally. Important
components of biomedical ethics, such as informed consent, algorithm fairness and
biases, intellectual property law, data privacy, safety, and transparency, should be
considered alongside the regulatory issues of DTs [36]. Computer modeling and
simulation can be adopted from other industries to aid in various stages of medical
device development, testing, clinical evaluations, and failure analysis, leading to cost
reduction [37].
Recently, the FDA has published reports to showcase how modeling and simula-
tion can be used for scientific research and regulatory decision-making [38]. FDA also
published guidance providing recommendations to assess the Credibility of Compu-
tational Modeling and Simulation in Medical Device Submissions [39]. Virtual clin-
ical trials represent a significant shift in how clinical research is conducted. By
leveraging DT platforms, researchers can gather large data more quickly and from a
broader population base. Simulations and virtual environments offer a unique oppor-
tunity for policymakers to test the implications of healthcare policies in a controlled,
risk-free setting [40]. By modeling the outcomes of proposed policies, stakeholders
can anticipate their effects and refine them before implementation, ensuring that new
regulations are both effective and efficient.

3 Health Professional Training Through Simulation

and Digital Twin Technology

Over the past two decades, there has been a significant expansion in clinical training
for healthcare professionals through various simulation technologies. These tech-
nologies range from task trainers, mannequins to more advanced Virtual Reality
(VR) and Augmented Reality (AR) systems. The learners include the multidisci-
plinary clinical team (physicians, nurses, pharmacists, and respiratory therapists).
While traditional task trainers have been around for many years, modern technology
has allowed for the expansion and sophistication of simulation training programs.
These programs now cover a wide array of skills, from procedural skills to non-
technical skills like communication, compassion, and empathy. Numerous studies
have shown that simulation-based training is associated with improved care processes
and better patient outcomes [41].
258 Y. Dong et al.

3.1 Computer-Driven Mannequins and Task Trainers

in Medical Training and Assessment

Modern mannequins used in medical training have evolved to become highly sophis-
ticated, computer-driven devices. These mannequins can be controlled remotely or
in-room, offering a range of clinical scenarios for trainees to practice clinical skills
and teamwork [42, 43]. Those mannequins provide invaluable hands-on experience
for clinicians in a controlled, risk-free environment for various specialties and disci-
plines [41, 44, 45]. The COVID-19 pandemic has forced the healthcare industry to
adopt distance simulation using various technologies to find alternative approaches
for traditional onsite clinical training [46, 47]. Simulation has been used for skill
assessment and evaluation in many disciplines [48–50]. The use of VR and AR
in healthcare training is growing rapidly to offer immersive learning experiences
for learners that traditional methods cannot match [51–53]. Game-based training
modules have also gained popularity, providing an engaging way for healthcare
providers to hone their skills [54]. Although the cost of hardware can be a limiting
factor, the benefits, such as improved skill retention and real-world applicability,
often outweigh the initial investment.

3.2 Research Potential: Investigating System Vulnerabilities

Simulation technologies not only provide clinical skill training but also serve as
research tools for human factor analysis. They offer a reproducible clinical environ-
ment where the interaction of system factors (clinician, technology, and workflow)
can be investigated independently of patient factors [55, 56]. Researchers can use
these technologies to investigate system vulnerabilities, test new workflows, new
technology innovations, and even simulate the impact of potential policy changes
before clinical implementation [57–60]. This simulation-based research extends the
utility of simulation technologies as integral tools for overall healthcare system
improvement [61].

4 Hospital Operations and Management

4.1 Leverage Data Insights from EMR, IoT, AI

The past 15 years have seen a data revolution in health care fueled by electronic
medical records’ (EMRs) implementation and adoption. This shift from paper
to digital documentation has unlocked a wealth of patient data, improving team
communication and clinical decision support, informing disease treatment plans,
and tracking outcomes. Further, the integration of EMRs with the IoT and Radio
Evolution of Simulation and Digital Twin in Health Care: From … 259

Frequency Identification (RFID) technologies creates a connected ecosystem for

real-time patient flow monitoring and streamlined inventory management benefits
from this data influx [62–64].
By collecting and analyzing this data, healthcare professionals can leverage
predictive modeling to anticipate patient needs and prescriptive analytics to design
personalize treatment plans, ultimately resulting in more informed decisions,
improved patient care, and efficient healthcare operations [65]. Recent advancements
in AI, computer vision, and Large Language Model (LLM) are opening up a new era
in the healthcare innovation using simulation and DT [66, 67]. These technologies
extend the capabilities of traditional data collection and analysis, offering a more
holistic view of healthcare delivery from outpatient settings to inpatient care, and
even post-discharge home care.

4.2 Operation Management and Resource Allocation

Healthcare operations’ management aims to improve the efficiency of business

processes within healthcare facilities to reduce overcrowding, waiting times, and
delays. One of the main challenges in healthcare operations is managing the unpre-
dictability caused by variations in patient demand, staffing capacity, and resource
availability. For instance, changes in the number of patients or types of cases can
affect staff schedules, patient flow, and patient room utilization. According to a rapid
literature review of papers published in 2002–2022, the use of DTs for healthcare
systems management is an emerging topic [68]. With availability of real-time patient
and hospital operation data, healthcare managers can make informed decisions about
staff deployment based on patient needs and workflow efficiencies. Similarly, data
analytics can be used to optimize inventory levels, ensuring that medical supplies
are ordered and utilized most effectively, thereby reducing waste and costs. The key
advantage of DTs in health care is the ability to leverage real-time data to model
complex systems and processes that involve many interdependent variables [69].
This allows for dynamic, evidence-based decision-making by integrating a large
amount of heterogeneous data and real-time data queries to achieve better resource
allocation at the hospital level [70].
Trauma centers, emergency departments, and ICUs are systems whose processes
are subject to large variability and are very time-sensitive, and thus, have attracted
substantial attention for process improvement and patient safety assurance using
DTs. For instance, a trauma DT is used to digitize and support the process of severe
trauma management, considering it as a physical asset that is mirrored by two DTs
[71]. In Augusto et al. 2018, a DT of an emergency unit was developed to optimize the
pathway of patient care in the unit. The system accounts for various arrival processes
to account for massive arrivals in case of a crisis and determine the best avail-
able leverages to optimize the operations of the system [72]. DTs of ICU processes
are used to identify inefficiencies in patient flows, optimize patient care by clinical
staff at the enterprise level, and use remote monitoring to detect process faults and
260 Y. Dong et al.

anomalies [68, 73]. Healthcare leaders can shift from reactive decisions to proactive
optimization based on data-driven insights from DTs. Hospital-level model by using
predictive decision support model that employs real-time service data is drawn from
the systems and devices [74]. Their model enables assessing the efficiency of existing
healthcare delivery systems and evaluating the impact of changes in services without
disrupting the daily activities of the hospital. Along the same line, Karakra et al. 2020
developed discrete event simulation and DTs through a system called HospiT’Win
that allows for tracking the pathways of patients inside the healthcare organization to
manage growing demand and decrease waiting times [69]. Rodriguez-Aguilar et al.
2020 proposed a digital healthcare system initiative through multi-paradigm simu-
lation [75]. Computer modeling has also been used to simulate infectious disease
transmission dynamics, optimize the vaccination strategy, and test public policies
before clinical implementation [76–78].

5 Challenges and Opportunities: Ushering a New Era

in Healthcare Simulation and Digital Twins

The adoption of simulation and DT technologies in health care is not without its chal-
lenges [79, 80]. One of the most pressing issues is the need for accurate and complete
data. Incomplete or erroneous data can significantly impact the effectiveness of these
technologies in both training and real-world applications. Also, modeling complex
biological systems involves numerous variables and nonlinear interactions, making
it challenging. Regulatory hurdles, such as compliance with healthcare standards
and data protection laws, further complicate the adoption process. Ethical consider-
ations are paramount when using patient data in the DTs, especially when it comes
to patient and clinician data privacy and confidentiality. The collection and use of
patient data must adhere to strict ethical guidelines to ensure that individual privacy
is respected. Moreover, there is a risk of bias in data collection and analysis, which
could inadvertently lead to unequal healthcare delivery. Addressing these ethical
and bias concerns is crucial for the responsible deployment of these technologies.
Data sharing and interoperability present another set of challenges. Different health-
care systems often use different data formats and standards, making integration a
complex task. The lack of interoperability can hinder the seamless exchange of
information, thereby limiting the effectiveness of simulation and DTs technologies
in a multi-system environment [81].
For simulation and DT tools to be effectively integrated into health care, there
needs to be a set of clearly defined performance metrics for systems evaluation. These
metrics should measure impacts of simulation and DTs across various domains,
including disease outcomes, patient outcomes, system outcomes, and return on
investment. The development of such comprehensive evaluation metrics will enable
stakeholders to assess the effectiveness of DT technologies objectively.
Evolution of Simulation and Digital Twin in Health Care: From … 261

Despite these challenges, there are numerous opportunities for innovation. For
instance, multiscale simulation is a powerful tool for understanding complex biolog-
ical systems and healthcare processes. By simulating the behavior of individual
molecules, cells, and tissues, multiscale simulations can provide insights into how
diseases develop and how drugs work. Multiscale simulations are being used to
understand the molecular basis of diseases, such as Alzheimer’s disease and drug
development [82, 83]. Meanwhile, DTs can be used to track the progress of patients,
monitor their vital signs, and predict when they may need medical attention. DTs have
been used to monitor cardiac and cancer patients for personalized treatment planning
[84, 85]. Additionally, AI algorithms can be used to analyze the data generated by
simulations and DTs. This data can be used to train AI models to make predictions
and identify patterns. AI models can be used to improve the accuracy of diagnoses,
recommend treatments, and develop new drugs. The synergistic relationship between
multiscale simulation, DTs, and AI offers unprecedented capabilities in health care,
from real-time analytics to predictive and prescriptive modeling for disease preven-
tion, diagnosis, and management. The integration of multiscale simulation, DTs, and
AI is creating a new era of precision medicine and personalized health care. Data-
driven decision-making is revolutionizing disease management and patient care. By
leveraging predictive analytics, healthcare providers can anticipate disease progres-
sion and intervene earlier. This proactive approach improves patient outcomes and
reduces the burden on healthcare systems. Simulation and DTs will empower clin-
icians and operational managers to use real-time data to streamline workflows and
enhance the quality of care. By analyzing patterns in healthcare delivery, they can
identify bottlenecks and inefficiencies, leading to policies that align with the goals
of all stakeholders and provide safe, effective, efficient, and equitable patient care.
The multidisciplinary nature of this field allows for the convergence of health-
care professionals, engineers, data scientists, IT experts, AI specialists, etc. Such
collaboration can lead to innovative solutions that address the existing challenges.
Moreover, as these technologies become more integrated into health care, there will
be a growing need for workforce training and upskilling to ensure that healthcare
professionals can effectively leverage these advanced tools.
While technological advances in health care are promising, their successful adop-
tion hinges on robust change management, implementation science strategies, and
workforce upskilling. Crucially, comprehensive training is needed to equip health-
care professionals with the skills to leverage new digital capabilities effectively. This
includes technical competencies, data literacy, human–AI collaboration, cyberse-
curity, and ethics. Investing in change management, implementation science, and
strategic workforce development will be critical for healthcare organizations to
capitalize on transformative technological advances. Several major societies and
groups are active in related fields. The Society for Simulation in Healthcare (SSH)
is a global community of medical educators who use various technologies, such as
manikins and task trainers, to deliver educational interventions for skills and team-
work training [86]. The Winter Simulation Conference includes many engineers
working on modeling and simulation in various industries [87]. The Interagency
262 Y. Dong et al.

Modeling and Analysis Group (IMAG) is a government group of program offi-

cials from multiple federal government agencies supporting research funding for
modeling and analysis of biomedical, biological, and behavioral systems. IMAG is
focused on research across the biological continuum across different scales or levels
of resolution with modelers from multidisciplinary research communities [88]. The
Medical Device Innovation Consortium (MDIC) primarily focuses on device inno-
vation, using Computational Modeling and Simulation (CM&S) to reduce product
development costs, speed up time to market, and better serve patients with safe and
effective medical devices [37]. The Virtual Physiological Human Institute for Inte-
grative Biomedical Research, commonly known as the VPH Institute, is another
significant player in this field. Its mission is to ensure that the Virtual Physiological
Human is fully realized, universally adopted, and effectively used in research and
clinical settings [89]. The Europe Digital Twin in Healthcare (EDITH) project aims
to capitalize on the growing trend of interest in Digital Twins [90]. EDITH creates
a roadmap for future development, allowing stakeholders to exchange best prac-
tices, analyze ecosystems and data flows, and identify vulnerabilities. In addition,
groups are working to prepare data for modeling. Mobilizing Computable Biomed-
ical Knowledge is an international community focused on ensuring that biomedical
knowledge in computable form is findable, accessible, interoperable, and reusable. A
recent report from the National Academies of Sciences, Engineering, and Medicine
presented an intergraded research agenda to advance the field. DTs can be a critical
tool for decision-making base on the synergistic combination of models and data
[91].

6 Conclusion

The emergence of simulation and DT technologies marks the beginning of a new era in
health care. These technologies are practical tools that can improve disease manage-
ment, care planning, and resource allocation. While there are undoubtedly chal-
lenges, such as technical complexities, ethical and regulatory hurdles, overcoming
these obstacles will enable us to fully integrate simulation and DT technologies to
support healthcare system digital transformation that serves all patients with safety,
effectiveness, patient-centeredness, timeliness, efficiency, and equity.

References

1. Yong PL., Olsen L, Jm M (2010) Value in health care: accounting for cost, quality, safety,
outcomes, and innovation: workshop summary
2. Brandt AM (2012) A reader’s guide to 200 years of the New England journal of medicine. N
Engl J Med 366(1):1–7
3. Carayon P et al (2020) SEIPS 3.0: human-centered design of the patient journey for patient
safety. Appl Ergon 84:103033
Evolution of Simulation and Digital Twin in Health Care: From … 263

4. Wickramasinghe N et al (2007) Healthcare system of systems. IEEE Int Conf Syst Syst Eng
2007:1–6
5. Institute of Medicine Committee on Quality of Health Care (2000) In A., To Err is human:
building a safer health system, In: Kohn LT, Corrigan JM, Donaldson MS (eds) To Err is Human:
building a safer health system. National Academies Press, National Academy of Sciences,
Washington (DC)
6. Makary MA, Daniel M (2016) Medical error—the third leading cause of death in the US. Br
Med J 353
7. Blumenthal D et al (2020) Covid-19—Implications for the health care system. N Engl J Med
383(15):1483–1488
8. Kaye AD et al (2020) Economic impact of COVID-19 pandemic on healthcare facilities and
systems: international perspectives. Best Pract Res Clin Anaesthesiol 35:293–306
9. Domecq JP et al (2021) Outcomes of patients with coronavirus disease 2019 receiving organ
support therapies: the international viral infection and respiratory illness universal study
registry. Crit Care Med 49(3)
10. Society for simuation in healthacre: what is simulation? (2024) 22 Apirl 2013. Available from:
http://ssih.org/about-simulation
11. Cook DA, Erwin PJ, Triola MM (2010) Computerized virtual patients in health professions
education: a systematic review and meta-analysis. Acad Med 85(10):1589–1602
12. Digital Twin Consortium (2024) Available from: https://www.digitaltwinconsortium.org/faq/
13. Kritzinger W et al (2018) Digital Twin in manufacturing: a categorical literature review and
classification. IFAC-PapersOnLine 51(11):1016–1022
14. Cockrell C et al (2022) Generating synthetic data with a mechanism-based critical illness digital
twin: demonstration for post traumatic acute respiratory distress syndrome. bioRxiv, p 517524,
22 Nov 2022
15. SIEMENS (2024) What is a digital twin? 2024; Available from: https://www.sw.siemens.com/
en-US/technology/digital-twin/
16. Pesapane F et al (2022) Digital twins in radiology. J Clin Med 11(21)
17. Bjelland Ø et al (2022) Toward a digital twin for arthroscopic knee surgery: a systematic review.
IEEE Access 10:45029–45052
18. Hernandez-Boussard T et al (2021) Digital twins for predictive oncology will be a paradigm
shift for precision cancer care. Nat Med 27(12):2065–2066
19. Rivera LF et al (2019) Towards continuous monitoring in personalized healthcare through
digital twins. In: Conference of the centre for advanced studies on collaborative research
20. Vallée A (2023) Digital twin for healthcare systems. Frontiers Digital Health 5
21. Pool R (1992) The third branch of science debuts. Science 256:44–47
22. Vodovotz Y (2023) Towards systems immunology of critical illness at scale: from single cell
‘omics to digital twins. Trends Immunol 44(5):345–355
23. Lal A et al (2020) Development and verification of a digital twin patient model to predict
specific treatment response during the first 24 hours of sepsis. Crit Care Explor 2(11):e0249
24. Lal A et al (2020) Artificial intelligence and computer simulation models in critical illness.
World J Crit Care Med 9(2):13–19
25. Alquraishi M, Sorger PK (2021) Differentiable biology: using deep learning for biophysics-
based and data-driven modeling of molecular mechanisms. Nat Methods 18:1169–1180
26. Yue R, Dutta A (2022) Computational systems biology in disease modeling and control, review
and perspectives. NPJ Syst Biol Appl 8
27. Eddy DM, Schlessinger L (2003) Validation of the archimedes diabetes model. Diabetes Care
26(11):3102–3110
28. Schlessinger L, Eddy DM (2002) Archimedes: a new model for simulating health care systems–
the mathematical formulation. J Biomed Inform 35(1):37–50
29. An G (2022) Specialty grand challenge: what it will take to cross the valley of death: transla-
tional systems biology, “True” precision medicine, medical digital twins, artificial intelligence
and in silico clinical trials. Frontiers Syst Biol 2
264 Y. Dong et al.

30. An G et al (2009) Agent-based models in translational systems biology. Wiley Interdiscip Rev
Syst Biol Med 1(2):159–171
31. McDaniel M et al (2019) A whole-body mathematical model of sepsis progression and treatment
designed in the BioGears physiology engine. Front Physiol 10
32. Bray A et al (2019) Pulse physiology engine: an open-source software platform for computa-
tional modeling of human medical simulation. SN Compr Clin Med 1:362–377
33. Rovati L et al (2024) Development and usability testing of a patient digital twin for critical
care education: a mixed methods study. Front Med 10
34. Pappalardo F et al (2019) In silico clinical trials: concepts and early adoptions. Briefings Bioinf
35. Viceconti M et al (2021) In silico trials: verification, validation and uncertainty quantification of
predictive models used in the regulatory evaluation of biomedical products. Methods 185:120–
127
36. Lal A et al (2022) Regulatory oversight and ethical concerns surrounding software as medical
device (SaMD) and digital twin technology in healthcare. Ann Transl Med 10(18):950
37. Landscape report and industry survey on the use of computational modeling and simulation in
medical device development (2023). Available from: https://mdic.org/resource/cmslandscape
report/#download_form
38. Successes and opportunities in modeling and simulation for FDA (2022)
39. Guidance document (2023) Assessing the credibility of computational modeling and simulation
in medical device submissions, U.S.F.D. Administration, Editor
40. Romero-Brufau S et al (2021) Public health impact of delaying second dose of BNT162b2 or
mRNA-1273 covid-19 vaccine: simulation agent based modeling study. BMJ 373:n1087
41. Cook DA et al (2011) Technology-enhanced simulation for health professions education: a
systematic review and meta-analysis. JAMA 306(9):978–988
42. Cooper JB, Taqueti VR (2004) A brief history of the development of mannequin simulators
for clinical education and training. Qual Saf Health Care 13:i11–i18
43. Gaba DM (2004) The future vision of simulation in health care. Qual Saf Health Care 13:i10–i12
44. Schmidt E et al (2013) Simulation exercises as a patient safety strategy a systematic review.
Ann Intern Med 158(5_Part_2):426–432
45. McGaghie WC et al (2014) A critical review of simulation-based mastery learning with
translational outcomes. Med Educ 48:375–385
46. Buléon C et al (2022) The state of distance healthcare simulation during the COVID-19
pandemic: results of an international survey. Adv Simul 7(1):10
47. Flurin L et al (2022) International virtual simulation education in critical care during COVID-19
pandemic: preliminary description of the virtual checklist for early recognition and treatment
of acute illness and iNjury program. Simul Healthc 17(3):205–207
48. Ryall T, Judd BK, Gordon CJ (2016) Simulation-based assessments in health professional
education: a systematic review. J Mult Healthc 9:69–82
49. Boulet JR, Murray DJ (2010) Simulation-based assessment in anesthesiology: requirements
for practical implementation. Anesthesiology 112(4):1041–1052
50. Hayden JK et al (2014) The NCSBN national simulation study: a longitudinal, randomized,
controlled study replacing clinical hours with simulation in Prelicensure nursing education. J
Nurs Regul 5(2, Supplement):S3–S40
51. Kononowicz AA et al (2019) Virtual patient simulations in health professions education:
systematic review and meta-analysis by the digital health education collaboration. J Med Int
Res 21
52. Foronda CL et al (2024) A comparison of virtual reality to traditional simulation in health
professions education: a systematic review. Simul Healthc 19(1s):S90-s97
53. Woodall WJ et al (2024) Does extended reality simulation improve surgical/procedural learning
and patient outcomes when compared with standard training methods?: A systematic review.
Simul Healthc J Soc Simul Healthc 19(1S):S98–S111
54. Gentry SV et al (2019) Serious gaming and gamification education in health professions:
systematic review. J Med Internet Res 21(3):e12994
Evolution of Simulation and Digital Twin in Health Care: From … 265

55. Small SD (2007) Simulation applications for human factors and systems evaluation. Anesthe-
siol Clin 25(2):237–259
56. Deutsch ES et al (2016) Leveraging health care simulation technology for human factors
research: closing the gap between lab and bedside. Hum Factors 58(7):1082–1095
57. Ahmed A et al (2011) The effect of two different electronic health record user interfaces
on intensive care provider task load, errors of cognition, and performance. Crit Care Med
39(7):1626–1634
58. Arriaga AF et al (2013) Simulation-based trial of surgical-crisis checklists. N Engl J Med
368(3):246–253
59. Patterson MD et al (2013) In situ simulation: detection of safety threats and teamwork training
in a high risk emergency department. BMJ Qual Saf 22(6):468–477
60. Yeh VJ et al (2024) Using mobile and remote simulation as a research methodology for health
care delivery research. Simul Healthc
61. Goldshtein D et al (2020) In situ simulation and its effects on patient outcomes: a systematic
review. BMJ Simul Technol Enhanc Learn 6(1):3–9
62. Asamoah DA et al (2016) RFID-based information visibility for hospital operations: exploring
its positive effects using discrete event simulation. Health Care Manag Sci 21:305–316
63. Haddara M, Staaby A (2018) RFID applications and adoptions in healthcare: a review on patient
safety. Procedia Comput Sci 138:80–88
64. Abkari SE, Jilbab A, Mhamdi JE (2020) RFID medication management system in hospitals.
Int J Online Biomed Eng 16:155–168
65. Sutton RT et al (2020) An overview of clinical decision support systems: benefits, risks, and
strategies for success. NPJ Digital Med 3
66. Cristofoletti M et al (2023) Towards a computer vision-based approach for digital twin
implementation. Int Conf Intell Metaverse Technol Appl (iMETA) 2023:1–6
67. Wang Y et al (2024) TWIN-GPT: digital twins for clinical trials via large language model
68. Elkefi S, Asan O (2022) Digital twins for managing health care systems: rapid literature review.
J Med Internet Res 24(8):e37641
69. Karakra A et al (2019) HospiT’Win: a predictive simulation-based digital twin for patients
pathways in hospital. IEEE EMBS Int Conf Biomed Health Informatics (BHI) 2019:1–4
70. Peng Y et al (2020) Digital twin hospital buildings: an exemplary case study through continuous
lifecycle integration. Adv Civil Eng 2020:1–13
71. Croatti A et al (2020) On the integration of agents and digital twins in healthcare. J Med Syst
44(9):161
72. Augusto V, Murgier M, Viallon A (2018) A modelling and simulation framework for intelligent
control of emergency units in the case of major crisis. Winter Simul Conf (WSC) 2018:2495–
2506
73. Zhong X et al (2022) A multidisciplinary approach to the development of digital twin models
of critical care delivery in intensive care units. Int J Prod Res 60(13):4197–4213
74. Karakra A et al (2018) Pervasive computing integrated discrete event simulation for a hospital
digital twin. In: 2018 IEEE/ACS 15th international conference on computer systems and
applications (AICCSA), p 1–6
75. Rodríguez-Aguilar R, Marmolejo-Saucedo JA (2020) Conceptual framework of digital health
public emergency system: digital twins and multiparadigm simulation. EAI Endorsed Trans
Pervasive Health Technol 6:e3
76. Currie CSM et al (2020) How simulation modelling can help reduce the impact of COVID-19.
J Simul 14:83–97
77. Romero-Brufau S et al (2021) Public health impact of delaying second dose of BNT162b2 or
mRNA-1273 covid-19 vaccine: simulation agent based modeling study. The BMJ 373
78. Zhang T et al (2023) Data-driven modeling and analysis for COVID-19 pandemic hospital beds
planning. IEEE Trans Autom Sci Eng 20:1551–1564
79. Fackler J, Hankin J, Young T (2012) Why healthcare professionals are slow to adopt modeling
and simulation. In: Proceedings of the 2012 winter simulation conference (WSC)
266 Y. Dong et al.

80. Combs CD, Combs PF (2019) Emerging roles of virtual patients in the age of AI. AMA J
Ethics 21(2):E153-159
81. Jacoby M, Usländer T (2020) Digital twin and internet of things—Current standards landscape.
Appl Sci
82. Stefanovski L et al (2021) Scales in Alzheimer’s disease: biological framework for brain
simulation with the virtual brain. Frontiers Neuroinformatics 15
83. Joshi A et al (2020) Opportunities for multiscale computational modelling of serotonergic drug
effects in Alzheimer’s disease. Neuropharmacology 174:108118
84. Coorey G et al (2022) The health digital twin to tackle cardiovascular disease-a review of an
emerging interdisciplinary field. NPJ Digit Med 5(1):126
85. Stahlberg EA et al (2022) Exploring approaches for predictive cancer patient digital twins:
Opportunities for collaboration and innovation. Front Digit Health 4:1007784
86. Society for Simulation in Healthcare (SSH) Council for accreditation of healthcare simulation
programs accreditation standards
87. 2024 Winter Simulation Conference (2024). Available from: https://meetings.informs.org/wor
dpress/wsc2024/
88. IMAG: Interagency Modeling and Analysis Group (2024)
89. Virtual Physiological Human Institute (2024). Available from: https://www.vph-institute.org/
90. Europe Digital Twin in Healthcare (2024). Available from: https://www.digitaleurope.org/pro
jects/ecosystem-digital-twins-in-healthcare-edith/
91. National Academies of Sciences, Engineering, and Medicine. Foundational Research Gaps and
Future Directions for Digital Twins (2024). Available from: https://www.nationalacademies.
org/our-work/foundational-research-gaps-and-future-directions-for-digital-twins
Digital Twins for Sustainable
Semiconductor Manufacturing

Ala Moradian

Abstract This chapter explores the applications of Digital Twins in smart semi-
conductor manufacturing, highlighting their potential to drive a more sustain-
able industry. Digital Twins enable advanced monitoring of chemical and energy
consumption, as well as other environmental impacts within semiconductor fabrica-
tion processes. The chapter reviews various initiatives and applications that utilize
Digital Twins for smart monitoring and provides guidance on extending these efforts
to further optimize semiconductor manufacturing. This includes enhancing device
performance, yield, and cost of ownership, while minimizing environmental impact.
The chapter concludes that achieving truly smart manufacturing requires a holistic
approach focused on optimizing specific equipment design and processes (such as
deposition, etch, lithography, chemical mechanical planarization) and fostering cohe-
sive collaboration among stakeholders from academia and industry. Addressing the
gaps in understanding physics, chemistry, sensor technology, software infrastructure,
data security, and establishing universal standards and protocols for data sharing
and integration are essential for realizing the full potential of Digital Twins in the
semiconductor industry.

Keywords Digital twins · Semiconductor manufacturing · Sustainability · Smart

manufacturing · Simulation · Modeling

1 Introduction

With the arrival of digital transformation innovations such as, Generative AI1 and
LLMs,2 Industry 4.0, autonomous vehicles, and Digital Twins the demand for semi-
conductor chip manufacturing continues to grow. As a result, sustainability measures

1 Artificial Intelligence.
2 Large Language Models.

A. Moradian (B)
Applied Materials Inc, Santa Clara, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 267
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_12
268 A. Moradian

and reducing the environmental impact will become a key area of focus for manu-
facturers as well as governments and local energy regulators. To address this sustain-
ability challenge, manufacturers along the entire value chain need to work both indi-
vidually as organizations and collectively as an industry to define goals and agree on
actions. Major semiconductor manufacturers have begun advocating for the industry
to enable a smart and digital future while simultaneously reducing environmental
impacts.
In situ monitoring of energy, water and gas consumption of semiconductor manu-
facturing plants will be the first critical step. As the famous Lord Kelvin once
mentioned, what you cannot measure you cannot improve. By synchronizing data
from multiple sensors, and leveraging domain knowledge and models, detailed on-
demand reports of carbon footprint and related consumption parameters from both
the fab and sub-fab can be generated. In addition to the monitoring the consumptions,
Digital Twins of manufacturing equipment can enable what-if scenarios with virtual
wafer processing and estimating their environmental impact. Ultimately, manufac-
turing equipment in the fab and the auxiliary equipment in the subfab would need to
be enhanced with built-in intelligence to optimize process and operational conditions
for not only on-wafer process outcome but also the environmental impact of the whole
manufacturing flow of future semiconductor devices. This sustainability-driven goal
would require calibrated processes and hardware Digital Twins that can realistically
predict impact of any changes to the system, not only at the environmental impact
level, but also for the integration and performance of the semiconductor devices.

1.1 Digital Twins and Promises for Smart Manufacturing

Smart manufacturing represents a transformative shift in the industrial landscape,

harnessing the power of digital tools to optimize production processes. At its core,
this paradigm leverages the systematic collection of data streams from an array of
sensors, notably IoT3 devices, which are strategically integrated into manufacturing
environments. The linchpin of this revolution is the deployment of Digital Twins,
software products or models created baed on physics or data-driven algorithms. These
twins encompass a diverse array of technologies, from machine learning to hybrid and
heuristic models. They play a pivotal role in the analysis of real-time data, enabling
manufacturers to gain unprecedented insights for monitoring, diagnostics, process
enhancement, and asset control. The importance of this digital ecosystem cannot be
overstated, as it empowers businesses to enhance efficiency, reduce downtime, and
ultimately compete more effectively in an increasingly dynamic global market.
With the foundational understanding of smart manufacturing and the pivotal role
that Digital Twins play in optimizing industrial processes, it becomes imperative
to delve into the broader landscape of Digital Twins across diverse industries, with
an emphasis on manufacturing. By examining real-world use cases and exploring

3 Internet of Things.
Digital Twins for Sustainable Semiconductor Manufacturing 269

the evolving nature of Digital Twins, we will gain a comprehensive understanding

of their significance in manufacturing and the broader industrial context. In a later
section we will dive into the use of Digital Twins in semiconductor manufacturing
and the important use cases in addressing the carbon footprint, energy, water, and
chemical consumptions of the industry.

1.1.1 Definition of Digital Twin

The concept of Digital Twins was first proposed by Michael Grieves in 2002 in
the context of product lifecycle management (PLM) [1]. Since then, there has been
multiple definitions suggested among which the definition by NASA and Grieves
are more broadly accepted. NASA described the Digital Twin as follows: “A Digital
Twin is an integrated multiphysics, multiscale, probabilistic simulation of an as-built
vehicle or system that uses the best available physical models, sensor updates, fleet
history, etc., to mirror the life of its corresponding flying twin” [2]. Grieves proposed
that the basic Digital Twin model has three components, viz. physical entity in real
space, virtual entity in virtual space, and the networks of information data that tie
the physical and virtual entities or space together [3]. Digital Twin (DT) constructs a
virtual replica of a physical scenario to monitor the behavior of the ongoing process,
with condition monitoring, detection of anomalies, and prediction of future trends
within the process. Based on the literature, the definitions of the Digital Twin is (see
Fig. 1):
. the digital representation of a physical entity with possible data and models. The
term “data” refers to the data from all the processes obtained during the run time
and the system’s development phase [4];
. it always incorporates synchronization with its related physical entities [5];
. it is a simulation model for the working of the related physical entity or process
[6].

The Digital Twin Consortium (DTC) which promotes awareness, adoption, inter-
operability, and development of Digital Twin technology, defined Digital Twin as a
virtual representation of real-world entities and processes, synchronized at a specified
frequency and fidelity [7].

Fig. 1 Concept of Digital

Twin was coined in SME
management forum
completing the cycle: using
PLM information in the sales
and service function [1]
270 A. Moradian

Another definition is based on ISO 23247 [8], which described a Digital Twin
as a “fit-for-purpose digital representation of an observable manufacturing element
(OME) with synchronization between the OME and its digital representation”. OMEs
include personnel, equipment, materials, processes, facilities, environment, prod-
ucts. Figure 2 shows a diagram of Digital Twin in manufacturing based on this
definition. According to a report by National Academy of Sciences focused on the
foundational research gaps and future directions for Digital Twins, the definition
was proposed as [14]: a set of virtual information constructs that mimics the struc-
ture, context, and behavior of a natural, engineered, or social system (or system-of-
systems), is dynamically updated with data from its physical twin, has a predictive
capability, and informs decisions that realize value. The bidirectional interaction
between the virtual and the physical is central to the digital twin.
Feng et al. 2023 demonstrated an implementation of the ISO23247 for data
requirements for Digital Twins in additive manufacturing [8].
DTC introduced the Digital Twin Capabilities Periodic Table (CPT) which is an
architecture and technology agnostic requirements definition framework, see Fig. 3.
It is aimed at organizations who want to design, develop, deploy, and operate Digital
Twins based on use case capability requirements versus the features of technology
solutions [9].
Several systematic bibliometric analyses have been conducted in the litera-
ture targeting Digital Twin framework in the context of smart manufacturing. As
an example, Warke et al. 2021 surveyed evolution, background, and implementa-
tion of critical enabling technologies such as data-driven decision-making, machine
learning, artificial intelligence, and deep learning. In their study, Warke et al. 2021
leveraged Scopus and Web of Science databases from 2016–21 for their analysis [10].
Dalibor et al. 2022 presented a systematic mapping of 356 Digital Twins publications
to characterize the applications in different domains. 20 different application domains
were presented, but only 70% of the studied work corresponded to the manufacturing
industry. Those applications were related to monitoring and controlling either before
production, to improve the design process, or during product lifetime [11].

Fig. 2 Diagram of Digital Twin implementation based on ISO23247 [8]

Digital Twins for Sustainable Semiconductor Manufacturing 271

Fig. 3 Digital Twin capabilities periodic table proposed by DTC [9]

Figure 4 shows network analysis of co-occurrence of keywords used in Warke

et al. 2021 study from Scopus database.
Several national funding initiatives for Digital Twin development and proliferation
have emerged to address technology requirements, data structure, cyber-security
provisions, and capability development in diverse industrial sectors. One of such
examples is Change2Twin, a forward-looking initiative funded by EU since 2020
on a variety of industrial applications. This EU-funded project helps manufacturing

Fig. 4 Network analysis of co-occurrence of keywords in Scopus [10]

272 A. Moradian

subject matter experts (SMEs) in their digitization efforts to deploy Digital Twins.
They are a consortium of 18 partners from all over Europe and promote the idea
that the concept of Digital Twin is one of the big game-changers in manufacturing
allowing companies to significantly increase their global competitiveness [12].
With a value of $12.9 billion in 2022, the Digital Twin market is expected to
grow at a CAGR of 35% to 40% through 2030, according to industry analysts. This
market growth is primarily driven by the increasing adoption of enabling technologies
such as AI, enterprise internet of things (IoT) platforms, augmented reality (AR),
and virtual reality (VR) [13]. In a comprehensive report by National Academies
(funded by the Department of Defense—Air Force Office of Scientific Research
and Defense Advanced Research Projects Agency, the Department of Energy, the
National Institutes of Health, and the National Science Foundation) the opportunities
enabled by Digital Twin were defined and the foundational research and resources
needed to support the development of Digital Twin technologies were identified
[14]. The report presented critical future research priorities and an interdisciplinary
research agenda for the field, including how federal agencies and researchers across
domains can best collaborate.

1.2 Digital Twins for Environmental Impact

and Sustainability

Digital Twins play a pivotal role in enhancing sustainability efforts across various
industries. The ability to simulate and analyze operations in real-time allows busi-
nesses to identify and implement sustainable practices. As industries increasingly
prioritize environmental responsibility, Digital Twins emerge as powerful tools to
monitor, analyze, and minimize carbon footprints across diverse sectors. In manufac-
turing, these twins optimize energy consumption and production efficiency, reducing
carbon footprints. In agriculture, Digital Twins aid precision farming, minimizing
resource wastage. Smart buildings employ Digital Twins to optimize energy usage,
contributing to eco-friendly urban development. Below is a few examples.
Data from two enterprises was used as a case study in an industrial DevOps
research project called Titan [15], which was focused on methods and tools for inte-
grating and analyzing big data from IoT devices in industrial manufacturing and
used to estimate power consumption. The authors proposed measures that can be
implemented in an industrial DevOps analytics platform, the Titan Control Center
[16]. Henning et al. 2021 discussed two industrial pilot cases, where analyzing power
consumption data can serve the goals of reporting, optimization, fault detection, and
predictive maintenance. In a pilot implementation of a power consumption analytics
platform, they showed how measuring real-time data processing, multilevel moni-
toring, temporal aggregation, correlation, anomaly detection, forecasting, visual-
ization can be implemented with a microservice-based architecture (see Figure 5),
stream processing techniques, and the fog computing paradigm, see Fig. 6 [16].
Digital Twins for Sustainable Semiconductor Manufacturing 273

Fig. 5 Microservice-based pilot architecture of the Titan Control center for analyzing electrical
power consumption [16]

Seegrun et al. 2023 conducted a systematic literature review (SLR) through Scopus
and Web of Science. Their query retrieved 523 records on Scopus and 281 on Web of
Science (data extraction in October 2022). After the elimination of 249 duplicates,
the dataset contained 555 records. These publications were categorized according to
their industry focus.
From the search results, application of Digital Twins were allocated throughout a
product’s lifecycle and assessed in terms of their technological maturity and sustain-
ability scope. To analyze the state of the art of research on Digital Twins in the context
of sustainability and to frame directions for future research, a SLR was conducted
to address the following research questions (RQs):
. RQ1: What are the application scenarios of Digital Twins in the context of
sustainability in the manufacturing industry?
. RQ2: In which product lifecycle (PLC) phases are Digital Twins applied in the
context of sustainability?
. RQ3: What is the technology readiness level (TRL) of the Digital Twin
applications?
. RQ4: For the attainment of which sustainability objectives are Digital Twins
currently applied?
. RQ5: What are the key directions for future research in the manufacturing industry
for sustainability-focused applications of Digital Twins?
The authors divided a product lifecycle into three stages: Beginning-of-life (BoL),
Middle-of-life (MoL) as well as End-of-Life (EoL) which are in turn subdivided into
the following lifecycle phases:
. BoL: Planning (BoL1), product development (BoL2), production (BoL3)
274 A. Moradian

Fig. 6 Titan dashboard implemented in Grafana [16]

. MoL: Distribution (MoL1), usage (MoL2), service (MoL3)

. EoL: Reuse/remanufacturing (EoL1), recycling (EoL2), disposal (EoL3).
In terms of technological maturity of the implementations of Digital Twin identi-
fied in the research by Seegrun et al. 2023, five levels of technology readiness levels
(TRL) were defined, see Fig. 7:
. TRL 1: Basic principles observed (e.g., SLR, survey)
. TRL 2: Technology concept formulated (e.g., conceptual framework)
. TRL 3: Experimental proof of concept (e.g., validated conceptual framework)
. TRL 4: Technology validated in academic environment.
. TRL 5: Technology demonstrated in relevant environment [17].
Figure 8 shows the share of the addressed lifecycle phases within the literature
body. Most Digital Twin approaches proposed are primarily located in the BoL stage
or span across multiple lifecycle phases.
Digital Twins for Sustainable Semiconductor Manufacturing 275

Fig. 7 Distribution of
literature body according to
product lifecycle and
technology maturity level
[17]

Fig. 8 Application
potentials of Digital Twins
within the literature body
based on the stage in product
lifecycles [17]

XMPro inc provides a generic Digital Twin platform that can track real-time data,
provide decision support From multiple data sources [18]. In one example, the plat-
form was used to provide a virtual representation of the wind farm (see Figure 9)
synchronized at a high frequency and fidelity, which enables real -time monitoring,
analysis, and optimization. By integrating data from sensors and historical perfor-
mance, the Digital Twin can predict the wind farm’s behavior and identify potential
issues before they occur, enabling proactive measures to be taken. The Digital Twin
remote operations center provides a platform for remote decision-making, allowing
operators to monitor and predict changes without disrupting operations, resulting in
improved performance, reduced maintenance costs, and enhanced safety.
Abdune et al. 2023 proposed a data-driven methodology for integrating the energy
consumption model into Digital Twins using techniques such as segmentation and
regression. It relies on power absorption measurement of industrial equipment to
generate energy consumption related parameters to be fed into the DT model to
monitor the current operating condition of the physical system. A case study on an
industrial robot was used to validate and assess the performance of the approach in
a laboratory environment [20].
276 A. Moradian

Fig. 9 Wind farm Digital Twin dashboard developed in XMPro [19]

1.2.1 Categories of Digital Twins

Physics-based category

Physics-based modeling describes the behavior of a system using physics equa-

tions and first principles. For example, Asrai et al. (2018) provided a modeling
approach that can trace the energy flow to compute the energy consumed in milling
processes based on active mechanisms of energy conversion within a machine tool
[21]. Lv et al. (2017) analyzed the spindle acceleration energy consumption of
machine tool (CNC) lathes based on the computation of the moment of inertia for the
spindle drive system [22]. For many situations physics-based modeling methods are
challenging to implement in the real-world applications because these models require
the foundational understanding of the governing physics and accurate physical prop-
erties for materials, and domain knowledge which isdifficult to quantify or estimate.
In one example, a virtual machine tool to estimate the energy consumption during
machining condition was also presented [23]. A genetic algorithm was then used to
reduce energy consumption. Seow et al. (2013) proposed an approach to estimate
the impact of direct and indirect factors of manufacturing on energy consumption
through what-if scenario simulations in order to reduce energy [24].

Data-Based Category

Data-based models are also referred to as data models, empirical models, or artifi-
cial intelligence methods. This category of Digital Twins mostly depend on time-
series statistical analyses and machine learning techniques. It can be considered as
Digital Twins for Sustainable Semiconductor Manufacturing 277

statistical modeling based on systematic evaluation of experimental results. As an

example for energy consumption monitoring and forecast, Karanjkar et al. (2018)
proposed an IoT-driven Digital Twin for energy optimization of an automated printed
circuit board assembly line with various sensors, as well as an open-source software
platform for data fusion and the energy Digital Twin [25]. Data-driven techniques
do not involve explicitly modeling the underlying physical system. Artificial Neural
Network (ANN) is the most used algorithm; Support Vector Regression and Random
Forest are other much used models [26]. Some researchers used unsupervised deep
learning to select and extract features of energy consumption combined with super-
vised deep learning to predict energy of machine tools [27]. Another approach
proposed was based on sensors to monitor the energy to distinguish between cases
in which a large amount of machine data is available allowing machine learning
methods and those in which only basic process information are accessible [28].
Intel published a white paper on optimizing factory performance using Digital
Twins and listed some of their Automated Factory Solutions (Intel® AFS). Two
of the examples from these solutions include Intel® Factory Recon which enables
manufacturers to use game-like, immersive graphics capabilities to instantly visualize
their operations better than ever before—how they’re running now, how they ran in the
past, and how they might run in the future. This tool is expected to help with reducing
Mean Time to Repair (MTTR) for factory incidents. Another example is Factory
Optimizer which is an AI-based control layer for simulation optimization and offline
analysis. It’s a tool that allows engineers to quickly and easily test changes to factory
assets or conditions [29]. TSMC has implemented data models and machine learning
for quality management, such as auto defect classification (ADC) using Human-in-
the-loop (HITL) approach which leverages both human and machine intelligence to
create machine learning models [30]. Similarly, Micron uses AI to improve yield and
quality of memory device manufacturing [31].

Heuristic and Semi-empirical

Considering the complexity of many manufacturing processes, in particular in multi-

physics, multiscale (length and time), it is imperative to fill the gaps in the theory
and fundamental physics, chemistry, material science, etc. with empirical insights.
Heuristic models are generally experience-based methods which are used to reduce
the need for calculations pertaining to some less studied characteristics of the
problem, such as equipment size, performance, or process/operating conditions,
nucleation, crystal growth faceting, etching in large aspect-ratios, etc. Such models
are usually used to describe a phenomenological process where not all the details are
understood from the scientific perspective. As an example, Chen et al. 2023 used a
heuristic approach to study an automated smart manufacturing framework based on
Digital Twins and Blockchain for IIoT.4 The data used in the Digital Twins are all
from the cluster generated after blockchain authentication [32]. Rehman et al. 2023

4 Industrial Internet of Things.

278 A. Moradian

proposes a heuristic approach for cognitive Digital Twin technology. The author
proposed the heuristic approach as a feature selection tool to enhance the cognitive
capabilities of a Digital Twin throughout the product design phase of production. The
proposed approach was validated using the use-case of Power Transfer Unit (PTU)
production [33].

Hybrid-Based Category

Hybrid modeling refers to the situation where part of a model can be formulated on the
basis of first principles and part of the model has to be inferred from data because of
a lack of understanding of the mechanistic details [34]. Huang et al. 2022 proposed
a modeling framework for hybrid learning-based Digital Twin for manufacturing
process with the corresponding trial implementation. The authors constructed a data
processing procedure to contextualize metadata sources across the process chain,
and a modeling pipeline for the integration of production domain knowledge and
AI techniques [35]. Langlotz et al. 2022 published a concept for Digital Twins that
are modeled by physics-based and data-driven models. Furthermore, this concept is
validated with the help of a use case that controls an energy management system of
a model-scale factory by finding the optimal time for charging and using the battery
[36].
Reduced order models (ROM) or surrogate models are critical to enable real-time
Digital Twins and expedite the run-time of models. These methods are in partic-
ular critical for high-fidelity physics or hybrid models where solving massive set of
equations resulted from discretization of computational domain into thousands or
millions of grid point or mesh cells. Some of such methods are reviewed by Es-haghi
et al. 2024 [37].
Another important area of research for enabling predictive Digital Twins, in
particular for applications where the data is scarce (or sparse) are AI-assisted
physics models, sometimes referred to as physics informed machine learning (PIML),
physics-informed neural networks (PINN). Raissi et al. 2019 are among the pioneers
of the PINN method [38]. In addition, Graph Neural Net (GNN) in another technique
with increased popularity. Based on inherent advantages of graph-based structures
for representing unstructured data, graph neural network has been receiving more and
more attention in predicting flow fields defined on unstructured computing nodes.
Some of these methods are based on Scarselli et al. 2009 [39] and Sanchez-Gonzalez
et al. 2020 [40].

State-Based Category

State-based models are related to the segmentation of the system into different func-
tions and sections, each of which is typically represented by an average behavior
Digital Twins for Sustainable Semiconductor Manufacturing 279

or characteristics. For example, in the context of energy consumption and moni-

toring, at the machine level, the energy consumption of the various operating modes—
switch-off, start-up, waiting, idle, ready for processing, and processing—are typi-
cally examined here. An average energy usage is calculated for each operating mode
and processing step. The energy usage is then predicted by merging the average
consumption for each operating mode as well as the transition from one state to
another. Wang et al. (2019) used this Digital Twin approach to create an event-driven
online machine state decision for an energy-efficient manufacturing system [41]. Jia
et al. (2017) suggested a finite state machine to model energy consumption transient
state during machining using the Pareto principal to identify the important ones and
then constructing a state transition chart to measure their duration [42]. Others such
as Dietmair and Verl 2009 proposed modeling machine energy consumption based
on statistical discrete states [43].

1.2.2 Examples of Digital Twins Used for Sustainability Applications

The interTwin project is funded by the European Union Horizon Europe Program
(2023-25) aimed at designing and building a prototype of an interdisciplinary
Digital Twin engine (DTE), based on a co-designed Blueprint Architecture. The
mission of the Digital Twin engine is to simplify & accelerate the development of
complex application-specific Digital Twins that benefits researchers, business, and
civil society. The effort attempts to extend the technical capabilities of the European
Open Science Cloud with modeling & simulation tools, demonstrate data fusion with
complex modeling & prediction technologies among other things. The program has
described several deliverables, one of which is to support the climate change use
cases and the implementation of the related impact decision support tools. Some of
the use cases of the program include studying tropical storms change in response to
climate change—The goal is to provide notebooks for scientists and policy makers
for running analysis on tropical cyclones on future projection data. Other example
of use cases are wildfire risk assessment in response to climate change, and flood
early warning in coastal and inland regions [44].
Another European research project called Trustworthy virtual experiments and
Digital Twins (ViDiT) was started, funded by the European Union through the Euro-
pean Partnership on Metrology, co-financed from the EU’s Horizon Europe Research
and Innovation Program and by the Participating States.
ViDiT focuses on virtual experiments and Digital Twins as key enabling tech-
nologies to achieve and realize European strategic policies devoted to sustainability
and digitalization within the complex framework of Industry 4.0 and the European
Green Deal. The consortium of this three-year project consists of 21 institutions: 8
National metrology institutes, 2 research centers close to industry, 5 universities and
6 companies. The project is coordinated by Physikalisch-Technische Bundesanstalt
(PTB), based in Braunschweig, Germany [45].
Green material optimal selection (GMOS) in product design is a key issue for
realizing sustainable manufacturing. To improve the accuracy and efficiency for
280 A. Moradian

Fig. 10 Optimal selection of green material for manufacturing based on Digital Twins [46]

green material optimal selection in product, a new method driven by Digital Twin
was proposed by Xiang et al. 2019 [46]. In their method, first a high-fidelity Digital
Twin of the physical product was developed. Then the actual data (physical) and
the cyber (modeling) data was combined, and the performance of the green material
selection was evaluated. The team used a laptop design example to showcase the
methodology, see Fig. 10.
Tao et al. 2018 presented a five-dimensional model for Digital Twins. In Tao’s
model M_DT = (PE, VE, DD, Ss, CN), where PE refers to physical entity, VE
is virtual entity, DD is Digital Twin data, Ss is service for PE and VE, and CN is
connection among all PE, VE, and Ss. Xiang et al. 2019 proposed a variation of this
methods for GMOS where the 5D variation for Digital Twin model (5D-EDTM) is
divided into two parts M = (PE, VE, SE) + (DE, CE); one part consists of PE which
refers to evolutionary model for physical space, VE refers to evolutionary model for
virtual space, SE is the evolutionary model in service, DE stands for data evolutionary
model, CE refers to connection model among all dimensions. Therefore, 5D-EDTM
for GMOS can be divided into five models [6]. Implementation of the 5D-EDTM
driven GMOS method applied to GMOS is depicted in Fig. 11.
As described earlier, Henning et al. [16] used the Titan architecture to monitor
and analyze the power consumption for two pilot cases. Davila et al. 2023 performed
a similar analysis included other types of energy different from electricity (i.e., gas,
heat, compressed air) and relevant resources for sustainability, such as water and
raw materials [47]. Their study also considered the privacy concerns of the data from
different sources or stakeholders. Basically, they proposed a pre-processing data to
Digital Twins for Sustainable Semiconductor Manufacturing 281

Fig. 11 Digital Twin-based method for green material selection methodology, 5D-EDTM [46]
282 A. Moradian

obfuscate the protected information about the production process or the machines
used. There are many approaches to achieve privacy-preserving data sharing, for
example encryption, anonymization, perturbation, encoding or more complex data
synthesis (creating data synthetically). Choosing the appropriate technique depends
mainly on how the data is going to be used downstream, and what level of privacy,
or disclosure risk is desired.
The methodology proposed by Davila et al. 2023 consists of three levels for the
development of the Digital Twin. In first level (connection and communication), a
process model of the manufacturing steps is created. The first step is defining the
system boundaries, starting with the raw materials, and ending with the finished
product. Afterward, all the required inputs and outputs of each of the machines
involved in the different production processes are determined. The different inputs
can be electricity, heat, gas, compressed air, or materials, within others. The output are
typically the processed materials, intermediate products and finally the product. All
the relevant measurement points within the process are determined, to install sensors
required for the measurement of the inputs. As a result of the process modeling,
the flow of resources into and out of the system boundaries can be measured and
monitored in real-time. The second level is concerned with data collection from
all the machines and different data sources (and types), such as sensors, machines,
production, supplier, and weather data. Finally, in the third level (i.e., cognitive, and
analytical level), the data is used to perform energy efficiency and sustainability
assessments and the costs and estimate the durations per product, see Fig. 12.
In an example, Davila et al. 2023 demonstrated how different processes that
involve a set of machines or production steps can be considered for an overall energy
efficiency and sustainability analysis.
Ma et al. 2022 proposed a sustainable smart manufacturing strategy based on
information management systems for energy-intensive industries (EIIs) from the
product lifecycle perspective. Their study encompasses a Digital Twin-driven opera-
tion mechanism, and an overall framework of big data cleansing and integration are
designed to describe sustainable smart manufacturing. By two examples, the authors
showed that the unit energy consumption and energy cost of production as well as the
costs of environmental protection can be decreased by implementing their proposed
strategy based on the ‘cradle-to-gate’ lifecycle data analysis [48].
Li et al. 2022 developed a data-driven hybrid petri-net (DDHPN) [49, 95] inspired
by both the state-based energy modeling and machine learning for establishing
the energy behavior meta-model. Gaussian kernel extreme learning machine was
proposed to fit the instantaneous firing speed of energy consumption continuous
transitions in DDHPN. DDHPN-based energy behavior model was driven by phys-
ical data under real-time working conditions, operating parameters, and production
load for generating a virtual data space of energy management, See Figs. 13, 14, 15,
and 16.
In the context of sustainable manufacturing, the Triple Bottom Line TBL supports
the concept of sustainability in industries by clarifying three dimensions: economic,
environmental, and social [50]. The concept of triple bottom line has been introduced
over a decade ago and is being discussed more recently in the context of Digital Twins
Digital Twins for Sustainable Semiconductor Manufacturing 283

Fig. 12 Sustainability Digital Twin architecture proposed by Davila et al. 2023 [47]

for sustainability. For example, Miehe et al. 2021 attempted to stimulate interdisci-
plinary scientific discourse by discussing the interpretation of sustainable production
and the role of Digital Twins [51]. They concluded that although sustainability repre-
sents a key factor of future production, it is not conclusively defined in order to be
technically applicable. Existing (bottom-up) approaches assessing the contribution
of DTs to sustainable production are therefore not considered comprehensively from
a contemporary perspective. In order to navigate the technical development of DTs in
an appropriate direction in the sense of sustainable production, a further (top-down)
perspective is necessary.
Chavez et al. 2022 identified suitable key performance indicators (KPIs) for a
Discrete Event Simulation (DES) model and evaluated the impact in a drone factory
in four scenarios that test final assembly processes. Based on this project (called
TWINGOALS and funded by European Institute of Innovation and Technology (EIT)
[5294]), the team integrated sustainability indicators in a simulation model of a drone
assembly cell at the Stena Industry Innovation Lab (SII-Lab), based at Chalmers
University of Technology in Gothenburg, Sweden. This study attempted to answer
284 A. Moradian

Fig. 13 Energy behavior modeling based on a data-driven hybrid petri-net by Li et al. 2022 [49]

Fig. 14 Energy management framework proposed by Li et al. 2022 for manufacturing [49]
Digital Twins for Sustainable Semiconductor Manufacturing 285

Fig. 15 IoT-based data collection for energy management system [49]

Fig. 16 Prototype of a Digital-Twin-based energy management system [49]

286 A. Moradian

Fig. 17 Incorporation of KPIs related to the triple bottom line principle in the product lifecycle
[53]

two questions first identifying the most suitable KPIs to assess the sustainability
of production systems through a DES (or Digital Twin), and second, identifying
improvement sustainability opportunities of production systems. See Fig. 17 for
an example of the selected KPIs to be used in a Digital Twin model with TBL
considerations.
In a study performed by Banerji et al. 2021, a real-time agent-based optimization
motion planning for energy consumption of a robotic cellular was investigated. The
research proposed a framework a qualitative analysis, and a quantitative comparison
based on energy consumption in a robotic experiment case [54].

2 Digital Twins for Semiconductor Manufacturing

Since the first publication by Michael Grieves in 2005 [1] there has been remarkable
interests in developing and deploying Digital Twins to enable and accelerate smart
manufacturing across many industries. Similarly, there has been extensive progress
and developments by academia [55]. It is not a surprise that there has been an ever-
growing excitement also by the semiconductor industry to partake in adopting the
technology; however, semiconductor manufacturing is still in the very early innings
of creating and capturing value from Digital Twins with proven efficacy.
In 2022, the US government implemented an industrial strategy to revitalize
domestic manufacturing, strengthen American supply chains, and accelerate the
industries of the future. The CHIPS and Science Act of 2022 [56], which buildsd on
this progress, making historic investments that would potentially poise U.S. workers,
local communities, and American businesses to win the race for the twenty-first
century. This act promises to strengthen American manufacturing, supply chains, and
national security, and invest in research and development, science, and technology
in the industries of tomorrow, including nanotechnology, clean energy, quantum
computing, and artificial intelligence.
Digital Twins for Sustainable Semiconductor Manufacturing 287

The Industrial Advisory Committee (IAC) advises the Secretary of Commerce on

the science and technology needs of the nation’s domestic microelectronics industry,
the national strategy on microelectronics research, the CHIPS research and devel-
opment programs and other advanced microelectronics activities authorized under
Section 9906 of the CHIPS Act, and opportunities for new public–private partner-
ships to advance the domestic microelectronics industry [90, 91]. The charter of this
working group has been to look at the long-term research needs of the semicon-
ductor industry. The working group will then need to understand where the gaps
are, and then suggest priorities to the IAC as to where the focus areas should be for
CHIPS funding and the National Semiconductor Technology Center (NSTC) that
provide the best opportunities to sustain US leadership in semiconductor innovation.
NSTC is a key component of the research and development program established by
President Biden’s CHIPS and Science Act. One of the recommendations from the
IAC on the R&D Gaps Working Group has been the development and use of Digital
Twins in semiconductor R&D and manufacturing. On May 6th, 2024, US Depart-
ment of Commerce announced $285 million Funding Opportunity for a Digital Twin
and Semiconductor CHIPS Manufacturing USA Institute [57]. CHIPS Digital Twins
Manufacturing USA Institute The objectives of the Institute include: Convene stake-
holders across the semiconductor production ecosystem; Improve the state of the
art in manufacturing-relevant digital twins; Significantly reduce cost for U.S. chip
development and manufacturing; Improve development cycle times of semiconductor
product innovation; Advance digital twin-enabled curricula for training a domestic
semiconductor workforce; and create a digital twin marketplace for industry to access
digital models.
Industries are shaping themselves toward the development of customized and
cost-effective processes to satisfy customer needs with the aid of a Digital Twins.
Such Digital Twin frameworks are expected to enable the user to monitor, simu-
late, control, optimize, and identify defects and trends within ongoing processes,
and reduces the chances of human prone errors. Digital Twins represent tremendous
opportunities to propel semiconductor manufacturing by accelerating prototyping,
testing, concept creating, process emulation, and knowledge sharing for talent devel-
opment. To help manage the increase in complexity, we must use every tool in our
toolbox. Recent advances in algorithms, modeling, and artificial intelligence, as well
as the integration of sensors and metrology, offer great promise to help overcome
these challenges. Predictive maintenance and diagnostics, chamber matching, accel-
erated process transfer, and equipment performance and productivity improvement
are some of the benefits of Digital Twins for equipment and processes. Following up
from a workshop by SEMI dedicated to Digital Twins for semiconductor manu-
facturing [58], The SEMI Smart Manufacturing Initiative published a summary
whitepaper with industry members presenting a comprehensive view of Digital Twin
technology in semiconductor manufacturing for achieving AI-driven autonomous
factories, covering industry definitions, taxonomy descriptions, and challenges in
development and deployment.
288 A. Moradian

2.1 Framework for Semiconductor Capital Equipment

Manufacturing

A successful Digital Twin framework for semiconductor manufacturing would

conceptually provide a flexible platform where different dimensions of the complex
ecosystem spanning from fleet-level down to on-wafer feature-level, and even mate-
rials engineering and atomistic-level, are captured and the interconnectedness among
the different levels are considered. In an ideal future implementation of such a frame-
work, the interconnectedness of processing reactors knobs and the quality and perfor-
mance of the device fabrication will be a two-way relationship where the virtual fab,
built upon a series of synchronized and high-fidelity Digital Twins, enables actionable
insights (or even autonomous action) for optimized process development in terms
of performance, sustainability, cost, time, or other optimization targets. In addition,
that future ideal implementation, would provide insights for managerial implica-
tions regarding how the process flow, tool, processing reactors can be selected or
re-arranged for overall optimum cost of ownership and revenue perspective.
In the realm of Digital Twins, frameworks and platforms specially tailored to
the intricate and demanding landscape of semiconductor manufacturing have been a
missing link. Recognizing this gap, Applied Materials proposed AppliedTwin™,
envisioned to address the unique challenges of the semiconductor industry.
AppliedTwin represents the entire integration flow of various processes in the semi-
conductor manufacturing chain. This holistic approach covers the entire spectrum
of Digital Twins application from fab-level models (FabTwin™) down to device
(DeviceTwin™), even extending its reach to encompass materials engineering. This
comprehensive platform can make significant strides in bridging the gap between the
Digital Twin concept and the complex needs of semiconductor manufacturing.
Per AppliedTwin™ framework, Digital Twin for semiconductor manufacturing
is a virtual representation of equipment and process built from the best available
models, along with sensors and metrology data, guided by domain knowledge, and
synchronized at a specified frequency with the physical equipment and process.
AppliedTwin predicts performance of the equipment and process with reasonable
fidelity and provides actionable insights.
Furthermore, Digital Twins for semiconductor equipment and processes are built
on foundations where critical behavior of an equipment and process is modeled and
synchronized with the physical/actual system through sensors and metrology data.
An essential element of the twin, however, is the domain knowledge which captures
process and hardware know-how, see Fig. 18.
Through its multiple classes of twins - such as ChamberTwin™, ProcessTwin™,
Tool Twin™, FleetTwin™ - AppliedTwin will provide a digital replica of the
actual system (hardware and process) and can enable virtual experimentation and
creating actionable insights, such as predicting the impact of reactor design or process
twinning knobs on feature-level topology and device performance on wafer, see
Fig. 19.
Digital Twins for Sustainable Semiconductor Manufacturing 289

Fig. 18 Foundations of Digital Twins for semiconductor manufacturing [59]

Fig. 19 Classifications and characteristics of Digital Twins for semiconductor manufacturing [60]

Some of the characteristics of a twin in the context of semiconductor manufac-

turing are connectivity, user segment, and user interface where the user can interact
with twin. the user interaction could of course consist of a graphical user intergace
(GUI), augmented/virtual/mixed reality (AR, VR, MR), or a textual and/or real-time
natural language communication enables by artificial intelligence. ChamberTwin,
for example can be either always in live communication with the physical chamber
or synchronized at a certain frequency. The twin could be also tailored to support
the design modification of the equipment or tuning a process recipe at the R&D
stage, ramp-up, or high-volume manufacturing. Such design modifications can be
enabled with the fast developing generative design algorithms and surrogate/ROM
model training and deployment methods. An example of ProcessTwin is described
by Sathiyanarayanan et al. [61].
It is projected that there will be a tremendous need for STEM and skilled workers
needed in near future in order to remain competitive in the semiconductor industry
[62–64]. For example, a study by Deloitte suggests that by 2030, more than one
290 A. Moradian

million additional skilled workers will be needed to meet demand in the semicon-
ductor industry [65]. To capture the domain knowledge (i.e., codifying the know-
how) and sustainable workforce development, EduTwin™ and SemiGuru™ repre-
sent all the educational aspects of the twins, for example virtual unit or integrated
processing for training new process engineers by providing an accurate replica of a
specific equipment or process where a virtual wafer can be processed, and results
can be analyzed. The virtual-processing capability developed based on the chamber,
processes, or fleet of twins would support democratization of the semiconductor tech-
nology knowledge. In multiple scopes, focused on different segmentsof the work-
force (,e.g., STEM students in high-schools, community colleges, universities, or
talent onboarding), or various levels (e.g., introductory to advanced), EduTwin and
SemiGuru can prepare the much-needed workforce and “bend the learning curve”
in training and onboarding technicians and engineers. Educational platforms based
on Digital Twins can prepare the much-needed workforce and "bend the curve"
in training and onboarding technicians and engineers. To properly safeguard intel-
lectual property and proprietary know-how, organizations should develop both an
internal and an external versions of such training platforms. For instance, EduTwin
could function as a more generic educational platform for training new industry
entrants, while SemiGuru could serve as an advanced educational platform containing
valuable details and know-how through twins with higher fidelity, accuracy, and
focus. Advancing digital twin-enabled curricula for training a domestic semicon-
ductor workforce is one of the objectives envisioned for the CHIPS Digital Twins
Manufacturing USA Institute.

2.1.1 Overcoming Key Challenges of Technology Inflections

Digital Twins and virtual fabrication clearly offer great promise, but there are chal-
lenges that need to be overcome to tap their full potential in the semiconductor
manufacturing environment. Semiconductor equipment and processes are highly
multiphysics and multiscale environments, with a diverse set of length and times
scales. It is extremely challenging to integrate these disparate length- and time-
scale models. In many cases, a complete understanding of the underlying physics
and chemistry for processes used in the semiconductor manufacturing is lacking.
Therefore, a hybrid-model, that not only rely on the physics but also incorporate
avaiable test data, are more desired. Another fundamental hurdle is having adequate
sensors providing real-time data that can be placed within the harsh environments
of the process chambers. The absence of appropriate sensors makes it difficult to
build accurate predictive models and algorithms that account for all the nuanced
interdependencies and nonlinear effects. Finally, there is the challenge of gathering
adequate amount of useful data to train data models and capturing adequate level of
physics to matching the models accurately with actual on-wafer parameters. This is
exactly why Digital Twins are built based on the best models and will need to be
revised and improved whenever more accurate data becomes available or there is a
major change in our fundamental understanding of the processes.
Digital Twins for Sustainable Semiconductor Manufacturing 291

Fig. 20 Schematic of Digital Twin for fab (FleetTwin) for heater jacker health and proper
installation [66]

In one example of Digital Twins application in semiconductor manufacturing,

Moradian et al. 2023 proposed using a digital representation of gas and chemical
delivery to processing chambers for fault detection and diagnostics for heater jackets
health and proper installation after preventative maintenance, see Fig. 20. This appli-
cation is categorized as a FleetTwin which covers gas delivery for a fleet of processing
chambers and can be incorporated in conjunction with other preventative mainte-
nance, diagnostics, or productivity Digital Twins within the manufacturing execution
system (MES) software infrastructure of a fabrication facility [66]. Another example
is EcoTwin™, which accuratly captures energy and chemical consumption as well as
carbon footprint of semiconductor manufacturing [92, 93(described in more details
later in this chapter). EcoTwin can be deployed as a chamber-level twin (with inclu-
sion of sub-fab consumptions related to the chamber), or extended all the way to the
fleet-level, where all the chambers and other fab and sub-fab equipment are taken
into account.

3 Sustainability in Semiconductor Manufacturing

The semiconductor industry is poised for significant growth, with expectations of

reaching a staggering one trillion dollars in revenue by the year 2030 [67]. However,
in light of the current energy consumption and carbon footprint associated with semi-
conductor manufacturing, it becomes imperative for the industry to make substantial
292 A. Moradian

Fig. 21 Carbon footprint of semiconductor manufacturing exceeding automotive industry [69]

advancements in resource efficiency and environmental sustainability. This trans-

formation is essential not only to seize the forthcoming business opportunities in
the next decade but also to ensure a greener and more sustainable future. Figure 21
compared carbon footprint of one of the major semiconductor manufacturers versus
a representative of automotive industry and suggests that the semiconductor manu-
facturing has already exceeded the automotive industry in terms of environmental
impact.
The current emissions from computing are almost 4% of the world total. This is
already more than emissions from the airline industry and are projected to rise steeply
over the next two decades. By 2040 emissions from computing alone will account
for more than half of the emissions budget to keep global warming below 1.5 °C.
Consequently, this growth in computing emissions is unsustainable. The emissions
from production of computing devices exceed the emissions from operating them,
so even if devices are more energy efficient producing more of them will make
the emissions problem worse [68]. The recent surge in the use of AI and large
language models, like ChatGPT, has highlighted the challenge of increased energy
consumption. ChatGPT consumes over half a million kilowatts of electricity each
day, an amount staggering enough to service about two hundred million requests.
ChatGPT’s daily power usage is nearly equal to 180,000 U.S. households, each using
about twenty-nine kilowatts.
Digital Twins for Sustainable Semiconductor Manufacturing 293

3.1 Energy Consumption and Carbon Footprint Challenges

Researchers led by Udit Gupta of Harvard University used publicly available sustain-
ability reports from companies including TSMC, Intel Corp. and Apple Inc. to
show that as computing becomes increasingly ubiquitous, “so does its environmental
impact.”
Information and computing technology is expected to account for as much as
20% of global energy demand by 2030, with hardware responsible for more of
that footprint than the operation of a system. “Chip manufacturing, as opposed to
hardware use and energy consumption, accounts for most of the carbon output,”.
Power use is more dramatic; TSMC’s annual electricity consumption is estimated
by Greenpeace at 4.8% of Taiwan’s entire usage, and more than that of the capital,
Taipei. Greenpeace says that will rise to 7.2% once commercial production comes
online of TSMC’s newest fabs that will shrink the process further from the current
leading-edge of 5 nm, or billions of a meter, to 3 nm chips [69].
It is also expected to consume 237 TWh of electricity globally in 2030, close to
Australia’s 2021 electricity consumption [70].
The Greenhouse Gas Protocol (GHG Protocol) offers companies different levels of
commitments to advance their sustainability journey. Emissions from semiconductor
device makers fall into different categories [71]:
Scope 1 emissions arise directly from fabs, primarily from process gases with
high global warming potential (GWP) that are used during wafer etching, chamber
cleaning, and other tasks; they can also come from high-GWP heat-transfer fluids
that may leak into the atmosphere when fabs use them in chillers. For example, NF3
has a global warming potential that is 17.000 times more potent than CO2 [72]. A
fraction of these potent gases are released to the atmosphere and directly contribute
to global warming.
Scope 2 emissions arise directly from purchased electricity, steam, heating, and
cooling equipment; the major sources include production tools and facilities/utilities.
Scope 3 emissions include all other indirect emissions in a company’s value chain;
upstream emissions are those generated by suppliers or their products, while down-
stream emissions are related to the usage of products containing semiconductors.
The number and quantity of materials used for IC manufacturing is steadily growing
with every technology node (e.g., Si wafers, bulk gases, gas precursors, minerals,
chemicals). The upstream production of these materials also leads to greenhouse gas
emissions that must be factored into the emission of the IC chip manufacturing [72].
Majority of the semiconductor fabrication emissions are from scope 1 process
gases and scope 2 electricity consumption of tools, see Fig. 22.
Semiconductor companies recognize that many of their customers have set aggres-
sive net-zero targets for their supply chain, see Fig. 23. To address these concerns,
some large semiconductor companies have begun to set ambitious sustainability
commitments.
Figure 24 shows the comparison of total carbon footprint (or emission) in kgCO2 /
Wafer for the different technology nodes from N28 to N5. The carbon footprint for
294 A. Moradian

Fig. 22 Majority of the semiconductor fab emissions are from scope 1 process gases and scope 2
electricity consumption of tools [71, 96]

full flow fabrication of an N5 device is almost two times larger than that of an N28
device technology [73].
Semiconductor companies are also committing to carbon reductions.
. Applied Materials is actively working with customers to look at how to reduce the
energy used in the fab during the manufacturing of the chips and aims to achieve
net-zero global emissions for scopes 1, 2 and 3 by 2040 (Applied Materials inc.,
2023, [87]).
. Lam Research has committed to achieve 100% renewable electricity in 2030, net-
zero operations (Scope 1 and 2) in 2040 and net-zero Scope 3 emissions by 2050
(Lam Research inc., 2023, [88]).
Digital Twins for Sustainable Semiconductor Manufacturing 295

Fig. 23 Many semiconductor companies and their customers have made aggressive net-zero
commitments [71]

Fig. 24 Evolution of total carbon footprint for different technology nodes [73]
296 A. Moradian

. ASM International expects to achieve net-zero emissions for scopes 1, 2 and 3 by

2035 (ASMI, 2021, [89])
. ASML expects to achieve net zero GHG emissions from product use at customers
by 2040 and emissions in operations (scope 1 and 2) by 2025 [74].

3.2 Standards Relevant to the Digital Twins

and Sustainability

SEMI S23 is the standard [75] that addresses concepts related to energy, utilities, and
materials measurement and use efficiency of semiconductor manufacturing equip-
ment . Semi S23 is intended to be a tool that can be used to analyze energy, utilities,
and materials usage. Other relevant standards include ISO 50001, which provides
guidance for energy management systems—Requirements with guidance for use,
as well as a framework for continues monitoring and improvement (i.e., Plan, Do,
Check, Act) [76]. Another relevant standard is ISO 20140 which provides directions
for automation systems and integration and evaluating energy efficiency as well
as other factors of manufacturing systems that influence the environment [77]. The
German Standards Institute (DIN) requires the Asset Administration Shell (AAS)—a
reference framework for Digital Twins [78]—to be suitable for containing sustain-
ability data and to provide it at the end of a product’s lifecycle for efficient disposal or
recycling. DIN provides a similar concept in the form of a lifecycle record for tech-
nical plants [51]. Further standardizations that specifically address isolated aspects
of sustainability (e.g., energy efficiency) in the field of Industry 4.0 and Digital Twins
is IEC 62,832–3 [79, 80].

3.3 Water Consumptions in a Semiconductor Fab

The largest use of water (about three-quarters) in a fab is process related, with much of
that being converted to ultra-pure water (UPW) needed for production itself, followed
by the facility scrubber and cooling tower (both about one-tenth), see Fig. 25. Fabs
typically have separate circuits for ultrapure water (UPW), which can be hot and
cold, and lower purity (LP) water. UPW generation is a complex, multistep process
that also consumes significant amounts of power. Most fabs have some level of UPW
reclamation, although rates vary widely among fabs and processes within a fab.
Many processes associated with water usage also contribute to the industry’s carbon
footprint through their consumption of energy. The latest update to S23, SEMI’s
guidance for energy use in the fab, assigns a value of 9 kWh/m3 to generate cold
UPW, and 92 kWh/m3 for hot UPW. S23 does not declare a value for wastewater
management.
Digital Twins for Sustainable Semiconductor Manufacturing 297

Fig. 25 An estimate of
water usage split in a
semiconductor fab [82]

Frost and Hua published a study where they tried to quantify spatio-temporal
impacts of the interaction of water scarcity and water use by the global semicon-
ductor manufacturing industry. Their analysis is useful as a benchmark for global
semiconductor industry water withdrawals and may assist OEMs in decisions about
supply chain sourcing. This could also guide semiconductor manufacturers in priori-
tizing locations and time periods to implement water-saving technologies or employ
less water intensive electricity sources. One highlight from this study was that in the
case of semiconductor manufacturing, the most efficient way to reduce overall manu-
facturing water withdrawals is through reduction in fab electricity use. Reductions
in electricity water use can also be achieved by using less water intensive sources of
electricity, such as solar PV and wind, which is especially important during seasons
of higher water scarcity [81].
The total water drawn from the network for different technology nodes is shown
in Fig. 26.

4 Digital Twins for Monitoring and Optimizing

Environmental Impacts in Semiconductor
Manufacturing

With the arrival of digital transformation innovations such as generative AI, industry
4.0, autonomous vehicles, and Digital Twins the demand for semiconductor chip
manufacturing continues to grow. As a result, sustainability measures and reducing
the environmental impact will become a key area of focus for manufacturers as well
as governments and local energy regulators. To address this sustainability challenge,
companies need to work individually as organizations and collectively as an industry
to define goals and agree on actions.
Efforts to address eco-efficiency in semiconductor manufacturing has been
ongoing for the past decade, for instance Higgs et al. 2012 reviewed historic trends in
energy use and environmental impacts of the finished products and identified some
of the key challenges for reducing these impacts going forward [83]. By creating
digital replicas of physical systems, Digital Twins can perform real-time monitoring,
298 A. Moradian

Fig. 26 The total water drawn from the network for different technology nodes [73]

analytics, and simulation of energy systems, infrastructure, and ecosystems. This can
aid in resource management, which is key to mitigating climate change [84].
Using Digital Twins for quantifying carbon footprint, energy, and chemical
consumptions of semiconductor manufacturing equipment was first implemented
in EcoTwin™ [85]. EcoTwin™ is a Digital Twin developed by Applied Materials
that enables on-demand energy and chemical consumption of semiconductor manu-
facturing equipment. By synchronizing data from multiple sensors, and leveraging
domain knowledge and models, detailed reports of carbon footprint and related
consumption parameters from both the fab and sub-fab are automatically generated.
EcoTwin Monitor provides transparency and actionable insights to identify opportu-
nities for sustainable solutions in product and process development [86]. Figure 27
illustrates convergence and aggregation of multiple data streams from subfab and
fab tools to portray a complete picture of consumptions and carbon footprint of the
manufacturing site. Such a digital thread can be extended across multiple fabs and
within as organization and across diverse geographical locations.
The implementation of EcoTwin is envisioned over a couple of distinct phases. In
its Monitor phase, EcoTwin provides transparency and actionable insights to iden-
tify opportunities for sustainable solutions in product and process development,
see Fig. 28. In the Explore phase, EcoTwin provides a myriad of tools aimed at
consumption predictions, what-if scenarios, and beyond. Ultimately, the manufac-
turers may adopt and implement opportunities for autonomous decision-making and
action-taking steps triggered by this system, at the chamber, tool, or fleet-level. This
Digital Twins for Sustainable Semiconductor Manufacturing 299

Fig. 27 Schematic of data streams aggregated to capture environment impact of manufacturing

process encompassing fab and subfab [85]

sustainability-driven goal would require calibrated processes and hardware Digital

Twins that can realistically predict impact of any changes to the system, not only at the
environmental impact level, but also for on-wafer process results. On-wafer process
results, such as deposition/etch uniformity, film quality, resistivity, critical character-
istic dimensions (CDs), etc., will remain as the primary priority for the industry. These
key performance indicators in turn determine device performance, yield, and cost of
ownership. Therefore, the calibrated and validated digital twins would need to eval-
uate and identify opportunities aimed at sustainability within the acceptable process
windows. In other words, despite the importance of sustainability goals, the tech-
nology node enablement remains as the highest priority. Fabs will rely on digital twins
as enablers for making trade-offs to ensure that the on-wafer process performance
is not affected while sustainability targets are met in high-volume manufacturing
(HVM) operations. The complexity of such Digital Twins, along with their neces-
sity for all fabs and throughout the entire semiconductor manufacturing value chain
requires the development of underlying technologies. This involves addressing gaps
in the fundamental understanding of physics, chemistry, material science, and alter-
native chemicals, and their environmental impact. This topic would be a highly
suitable research area for the CHIPS Digital Twin Manufacturing USA Institute.
Figure 29 shows a user-friendly illustration of consumptions for the system (tool-
level shown) with the sub-systems, e.g., chambers and subfab auxiliary/supporting
equipment. In Fig. 30 detailed consumption information for a chamber with its
subcomponents (e.g., heaters, plasma source) are shown, as an example.
All the processing recipes ran during a selected date range can be evaluated and
ranked based on environmental impact and their amounts of chemicals consump-
tion. In addition, user can select specific recipes (e.g., best known-method process
recipes for a desired technology application) and compared them in detail from
environmental impact perspective; see Fig. 31.
300 A. Moradian

Fig. 28 Overall capability snapshot of EcoTwin Monitor for semiconductor manufacturing carbon
footprint and consumption monitoring platform [86]

Fig. 29 EcoTwin Tool-level dashboard – user can monitor overall impact as well as diving into
each of the subcomponents in fab or subfab

Fig. 30 Chamber-level dashboard with detailed consumptions and carbon footprint contributions
Digital Twins for Sustainable Semiconductor Manufacturing 301

Fig. 31 Process recipe consumptions, direct electrical, carbon footprints, and volumetric consump-
tion of chemicals

5 Conclusion

The role of Digital Twin technology in semiconductor manufacturing is rapidly

evolving, offering a powerful framework for tackling the complexities and chal-
lenges inherent in the industry. As Digital Twins with diverse fidelity levels and
complexities, informed by both physics-based models and sensor data, gain traction
across various sectors, their adoption in semiconductor manufacturing stands out as
particularly promising.
Achieving truly smart manufacturing for semiconductor industry requires a
holistic approach focused on optimizing specific equipment design and processes
(such as deposition, etch, lithography, chemical mechanical planarization) and
fostering cohesive collaboration among stakeholders from both academia and
industry. Addressing the gaps in a) understanding the science, b) developing more
accurate and new sensor technology, c) creating scalable software infrastructure
with considerations for intellectual property and (IP) and data security, and d) estab-
lishing universal standards and protocols for data sharing and integration, are essen-
tial for realizing the full potential of Digital Twins in the semiconductor industry. In
an era marked by the proliferation of AI-enabled applications, the emergence of
large-language models, and the relentless pursuit of accelerated computing methods,
Digital Twins present a compelling solution to address the multifaceted demands
of research and development, ramp-up, and high-volume manufacturing in the
semiconductor realm.
Beyond these overarching advantages, Digital Twins offer a multitude of specific
benefits tailored to the semiconductor manufacturing process. From process recipe
optimization and reactor design co-optimization for advanced technology nodes to
enhancing productivity in process engineering and ensuring chamber-to-chamber
matching, the potential applications of Digital Twins are vast and varied.
302 A. Moradian

Moreover, Digital Twins pave the way for workforce development initia-
tives centered around virtual representations of manufacturing equipment, integra-
tion processes, and device design and verification. By leveraging Digital Twins
of virtual fabs, personnel can engage in immersive training experiences and gain
invaluable insights into the intricacies of semiconductor manufacturing.
In essence, the integration of Digital Twin technology holds the promise of
revolutionizing semiconductor manufacturing, offering a pathway to enhanced effi-
ciency, innovation, and competitiveness in an increasingly dynamic landscape. As
the industry continues to evolve, Digital Twins are poised to play an indispens-
able role in driving progress and unlocking new possibilities on the semiconductor
manufacturing frontier.

References

1. Grieves M (2005) Product lifecycle management: the new paradigm for enterprises. Int J Prod
Dev 2(1–2):71–84
2. Shafto M, Conroy M, Doyle R, Glaessgen E, Kemp C, LeMoigne J, Wang L (2010) Modeling,
simulation, information technology and processing roadmap, NASA
3. Grieves M (2015) Digital Twin: manufacturing excellence through virtual factory replication
4. Negri E, Fumagalli L, Macchi M (2017) A Review of the roles of Digital Twin in CPS-based
production systems. Proc Manuf 11:939–948
5. Tao F, Zhang H, Liu A, Nee AYC (2019) Digital Twin in industry: state-of-the-art. IEEE Trans
Ind Inf 15(4):2405–2415
6. Tao F, Zhang M, Liu Y, Nee A (2018) Digital Twin driven prognostics and health management
for complex equipment. CIRP Ann 67(1):169–172
7. DTC, DIgital Twin Consortium (2020) [Online]. Available: https://www.digitaltwinconsort
ium.org/hot-topics/the-definition-of-a-digital-twin/
8. Feng S, Jones A, Shao G (2023) ISO 23247–1 Automation systems and integration—Digital
twin framework for manufacturing—Overview and general principles, International organi-
zation for standardization. In: Data requirements for digital twins in additive manufacturing,
ASME manufacturing science and engineering conference, New Brunswick, NJ, US
9. Digital Twin Consortium (DTC), Capabilities Periodic Table (CPT) (2022) [Online]. Available:
https://www.digitaltwinconsortium.org/initiatives/capabilities-periodic-table/
10. Warke V, Kumar S, Bongale A, Kotecha K (2021) Sustainable development of smart
manufacturing driven by the Digital Twin framework: a statistical analysis. Sustainability
13:10139
11. Dalibor M, Jansen N, Rumpe B, Schmalzing D, Wachtmeister L, Wimmer M, Wortmann A
(2022) A Cross-domain systematic mapping study on software engineering for Digital Twins.
J Syst Softw 193:111361
12. Change2Twin Project, Change2Twin—Digital Twin for every manufacturing SME! (2020)
[Online]. Available: https://www.change2twin.eu/
13. Land I, Niatas J, Serughetti M (2023) What are Digital Twins? A primer on virtual models
(2023) [Online]. Available: https://www.synopsys.com/blogs/chip-design/digital-twins-sem
iconductor-industry.html
14. National Academies of Sciences, Engineering, and Medicine (2024) Foundational Research
Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press.
https://doi.org/10.17226/26894.
15. Henning S, Hasselbring W (2012) The Titan Control center for industrial devops analytics
research. Softw Impact 7:100050
Digital Twins for Sustainable Semiconductor Manufacturing 303

16. Henning S, Hasselbring W, Burmester H, Möbius A, Wojcieszak M (2021) Goals and measures
for analyzing power consumption data in manufacturing enterprises. J Data Inf Manag 3(1):65–
82
17. Seegrün A, Kruschke T, Mügge J, Hardinghaus L, Knauf T, Riedelsheimer T, Lindow K (2023)
Sustainable product lifecycle management with Digital Twins: a systematic literature review.
Proc CIRP 119:776–781
18. XMPro inc (2023) [Online]. Available: https://xmpro.com/platform/. Accessed 11 Nov 2023
19. Green G (2023) XMPRO-Digital Twin consortium 2023: a digital twin-based remote operations
center for wind farms. [Online]. Available: https://www.brighttalk.com/webcast/18347/592103
20. Abdoune F, Ragazzini L, Nouiri M, Negri E, Cardin O (2023) Toward Digital Twin for sustain-
able manufacturing: a data-driven approach for energy consumption behavior model generation.
Comput Ind 150:103949
21. Asrai RI, Newman ST, Nassehi A (2018) A mechanistic model of energy consumption in
milling. Int J Prod Res 56(1–2):642–659
22. Lv J, Tang R, Tang W, Liu Y, Zhang Y, Jia S (2017) An investigation into reducing the spindle
acceleration energy consumption of machine tools. J Clean Prod 143:794–803
23. Lee W, Kim SH, Park J, Min BK (2017) Simulation-based machining condition optimization
for machine tool energy consumption reduction. J Clean Prod 150:352–360
24. Seow Y, Rahimifard S (2011) A framework for modelling energy consumption within
manufacturing systems. CIRP J Manuf Sci Technol 4(1):258–264
25. Karanjkar N, Joglekar A, Mohanty S, Prabhu V, Raghunath D, Sundaresan R (2018) Digital
Twin for energy optimization in an SMT-PCB assembly line. In: 2018 international conference
on internet of things and intelligence system (IOTAIS)
26. Walther J, Weigold M (2021) A systematic review on predicting and forecasting the electrical
energy consumption in the manufacturing industry. Energies 14(4):968
27. He Y, Wu P, Li Y, Wang Y, Tao F, Wang Y (2020) A generic energy prediction model of machine
tools using deep learning algorithms. Appl Energy 275:115402
28. Sossenheimer J, Walther J, Fleddermann J, Abele E (2019) A sensor reduced machine learning
approach for condition-based energy monitoring for machine tools. Proc CIRP 81:570–575
29. Lee D, Meyer JS, Touloukian P, Story J, Schneider P, Sartini J (2023) White paper: how Intel®
automated factory solutions delivers tremendous benefits to semiconductor manufacturing
environments, Intel
30. TSMC—Machine Learning for Quality Management (2023) [Online]. Available: https://www.
tsmc.com/english/dedicatedFoundry/services/apm_intelligent_packaging_fab/intelligentFab_
automation
31. Smart sight: how Micron uses AI to enhance yield and quality (2023) [Online]. Avail-
able: https://www.micron.com/insight/smart-sight-how-micron-uses-ai-to-enhance-yield-and-
quality
32. Chen H, Jeremiah S, Lee C, Park J (2023) A Digital Twin-based heuristic multi-cooperation
scheduling framework for smart manufacturing in IIoT environment. Appl Sci 13(3):1440
33. Rehman A, Ahmed M, Begum S (2023) Cognitive Digital Twin in manufacturing: a heuristic
optimization approach. In: Artificial intelligence applications and innovations; AIAI 2023. IFIP
advances in information and communication technology, vol 676
34. Bhalode P, Metta N, Chen Y, Ierapetritou M (2020) Efficient data-based methodology for model
enhancement and flowsheet analyses for continuous pharmaceutical manufacturing. Comput
Aided Chem Eng 48(2020):127–132
35. Huang Z, Fey M, Liu C, Beysel E, Xu X, Brecher C (2023) Hybrid learning-based digital twin
for manufacturing process: modeling framework and implementation. Robot Comput-Integr
Manuf 82:102545
36. Langlotz P, Klar M, Yi L, Hussong M, Sousa FJ, Aurich JC (2022) Concept of hybrid modeled
digital twins and its application for an energy management of manufacturing systems. Proc
CIRP 112:549–554
37. Es-haghi MS, Anitescu C, Rabczuk T (2024) Methods for enabling real-time analysis in digital
twins: a literature review. Comput Struct 297:107342
304 A. Moradian

38. Raissi M, Perdikaris P, Karniadakis G (2019) Physics-informed neural networks: a deep learning
framework for solving forward and inverse problems involving nonlinear partial differential
equations. J Comput Phys 378:686–707
39. Scarselli F, Gori M, Tsoi AC, Hagenbuchner M, Monfardini G (2009) The graph neural network
model. IEEE Trans Neural Netw 20(1):61–80
40. Sanchez-Gonzalez A, Godwin J, Pfaff T, Ying R, Leskovec J, Battaglia PW (2020) Learning
to simulate complex physics with graph networks. In: Proceedings of the 37th international
conference on machine
41. Wang J, Huang Y, Chang Q, Li S (2019) Event-driven online machine state decision for energy-
efficient manufacturing system based on digital twin using max-plus Algebra. Sustainability
11(18):5036
42. Jia S, Tang R, Lv J, Yuan Q, Peng T (2017) Energy consumption modeling of machining
transient states based on finite state machine. Int J Adv Manuf Technol 88:2305–2320
43. Dietmair A, Verl A (2009) A generic energy consumption model for decision making and
energy efficiency optimisation in manufacturing. Int J Sustain Eng 2(2):123–133
44. interTwin Project, "interTwin," (2023) [Online]. Available: https://www.intertwin.eu/about-int
ertwin/
45. Schmelter S (2023) Trustworthy virtual experiments and digital twins (ViDiT)—Funded
[Online]. Available: https://www.vidit.ptb.de/home
46. Xiang F, Zhang Z, Zuo Y, Tao F (2019) Digital Twin driven green material optimal-selection
towards sustainable manufacturing. Proc CIRP 81:1290–1294
47. Davila R, Maria F, Schwark F, Dawel L, Pehlken A (2023) Sustainability Digital Twin: a tool
for the manufacturing industry. Proc CIRP 116:143–148
48. Ma S, Ding W, Liu Y, Ren S, Yang H (2022) Digital twin and big data-driven sustainable smart
manufacturing based on information management systems for energy-intensive industries. Appl
Energy 326:119986
49. Li H, Yang D, Cao H, Ge W, Chen E, Wen X, Li C (2022) Data-driven hybrid petri-net based
energy consumption behaviour modelling for digital twin of energy-efficient manufacturing
system. Energy 239:122178
50. Almström P, Andersson C, Öberg A, Hammersberg P, Kurdve M, Landström A, Shahbazi S,
Wiktorsson M, Windmark C (2017) Sustainable and resource efficient business performance
measurement systems—The handbook, Mölndal, Mölndal: Billes Tryckeri
51. Miehe R, Waltersmann L, Sauer A, Bauernhansl T (2021) Sustainable production and the role
of Digital Twins–basic reflections and perspectives. J Adv Manuf Process 3(2):e10078
52. TwinGoals (2020) EIT manufacturing and ‘digital twins’ improving tomorrow’s manufac-
turing. [Online]. Available: https://eit.europa.eu/news-events/news/eit-manufacturing-and-dig
ital-twins-improving-tomorrows-manufacturing
53. Chávez CAG, Bärring M, Frantzén M, Annepavar A, Gopalakrishnan D, Johansson B (2022)
Achieving sustainable manufacturing by embedding sustainability KPIs in Digital Twins. In:
Winter simulation conference, Singapore
54. Barenji AV, Liu X, Guo H, Li Z (2021) A Digital Twin-driven approach towards smart
manufacturing: reduced energy consumption for a robotic cell. Int J Comput Integr Manuf
34(7–8):844–859
55. Kapteyn M, Pretorius J, Willcox K (2021) A probabilistic graphical model foundation for
enabling predictive digital twins at scale. Nat Comput Sci 1(5):337–347
56. The White House—Statements (2022) Fact sheet: chips and science act will lower costs,
create jobs, strengthen supply chains, and counter China. [Online]. Available: https://www.
whitehouse.gov/briefing-room/statements-releases/2022/08/09/fact-sheet-chips-and-science-
act-will-lower-costs-create-jobs-strengthen-supply-chains-and-counter-china/
57. U.S. Department of Commerce (2024) CHIPS for America Announces $285 million funding
opportunity for a digital twin and semiconductor chips manufacturing usa institute, U.S.
Department Of Commerce, [Online]. Available: https://www.commerce.gov/news/press-rel
eases/2024/05/chips-america-announces-285-million-funding-opportunity-digital-twin
Digital Twins for Sustainable Semiconductor Manufacturing 305

58. da Silva M, Somani K (2024) Digital twins in semiconductor manufacturing—semi smart

manufacturing initiative, SEMI, San Jose, CA
59. Kelkar U (2023) Semiconductor equipment and processes need Digital Twins.
[Online]. Available: https://www.appliedmaterials.com/us/en/blog/blog-posts/semiconductor-
equipment-and-processes-need-digital-twins.html
60. Moradian A, Kelkar U (2023) SEMI semiconductor Digital Twin workshop, Milpitas, CA:
SEMI
61. Rajesh S, Sadanandam N, Kailash M, Gowdra Thippeswamy S, Abhra R, Phillip S, Yi X,
Yu L, Wei L (2024) Coupling reactor-scale and feature-scale simulations: ProcessTwin™ for
unit processes. In: 2024 8th IEEE electron devices technology & manufacturing conference
(EDTM), Bangalore, India
62. Johnson M (2023) USA today. [Online]. Available: https://www.usatoday.com/story/opinion/
2023/09/13/stem-workforce-shortage-semiconductor-manufacturing-race/70721967007/
63. Patel P (2023) U.S. Universities are building a new semiconductor workforce. [Online].
Available: https://spectrum.ieee.org/chips-act-workforce-development
64. Rosso D (2023) America faces significant shortage of tech workers in semiconductor industry
and throughout U.S. economy. [Online]. Available: https://www.semiconductors.org/america-
faces-significant-shortage-of-tech-workers-in-semiconductor-industry-and-throughout-u-s-
economy/
65. Weisz K, Stewart D, Simons C, Lewis T, Kulik B (2023) The global semiconductor talent
shortage
66. Moradian A, L’Heureux JO, Sheng S, Mahakali R, Ramanathan K, Zhang L, Kelkar UM,
Prabhu GB, Yuan Z, Oh J (2020) Parameter sensing and computer modeling for gas delivery
health monitoring. United States Patent US11768984B2
67. Dickerson G (2020) Applied materials CEO Keynote—SEMICON Wes: make possible a better
future
68. Vanderbauwhede W (2022) Frugal computing—On the need for low-carbon and sustainable
computing and the path towards zero-carbon computing. IAB workshop on environmental
impact of internet applications and systems
69. www.supplychainbrain.com, The chip industry has a problem with its giant carbon foot-
print (2021) [Online]. Available: https://www.supplychainbrain.com/articles/32910-the-chip-
industry-has-a-problem-with-its-giant-carbon-footprint
70. Pelé A-F (2023) Semiconductor manufacturing on the way to net zero, [Online]. Available:
https://www.eetimes.com/semiconductor-manufacturing-on-the-way-to-net-zero/
71. Göke S, Issler M, Liu D, Patel M, Spiller P (2022) Keeping the semicon-
ductor industry on the path to net zero, McKinsey & Company. [Online]. Avail-
able: https://www.mckinsey.com/industries/semiconductors/our-insights/keeping-the-semico
nductor-industry-on-the-path-to-net-zero
72. Ragnarsson L-Å, Rolin C, Shamuilia S, Parton E (2022) The green transition of the IC industry,
imec—SSTS Program, Leuven
73. imec-NetZero, "imec.netzero" (2023). [Online]. Available: https://netzero.imec-int.com/.
Accessed 10 Nov 2023]
74. ASML-Sustainability, "Our sustainability strategy" (2022) [Online]. Available: https://www.
asml.com/en/company/sustainability
75. SEMI Standards (2023) SEMI S23—Guide for conservation of energy, utilities and materials
used by semiconductor manufacturing equipment, vol Safety Guidelines, SEMI
76. ISO-50001 (2018) ISO 50001 Energy management systems—Requirements with guidance for
use
77. Ghita M, Siham B, Hicham M, Griguer H (2021) Digital Twins based LCA and ISO 20140 for
smart and sustainable manufacturing systems. In: Sustainable intelligent systems. advances in
sustainability science and technology, Singapore, Springer
78. Boss B et al (2020) Digital Twin and asset administration shell concepts and application in the
industrial internet and industrie 4.0. Plattform Ind 4:13–14
306 A. Moradian

79. IEC-62832–3 (2020) BS EN IEC 62832–3 industrial-process measurement, control and

automation. Digital factory framework. Institution British Standards
80. ISO-20140 (2018) ISO 20140 Automation systems and integration—Evaluating energy
efficiency and other factors of manufacturing systems that influence the environment
81. Frost K, Hua I (2019) Quantifying spatiotemporal impacts of the interaction of water
scarcity and water use by the global semiconductor manufacturing industry. Water Resour
Ind 22:100115
82. Davis S (2022) Water supply challenges for the semiconductor industry. [Online]. Available:
https://www.semiconductor-digest.com/water-supply-challenges-for-the-semiconductor-ind
ustry/
83. Higgs T, Brady T, Yao M (2012) Progress and challenges in EcoDesign of semiconductor
products. In: Design for innovative value towards a sustainable society, Dordrecht
84. Verdict.co.uk: Intelligence, GlobalData Thematic (2023) Digital twins: key to addressing
climate change. [Online]. Available: https://www.verdict.co.uk/digital-twins-combat-climate-
change/
85. Moradian A (2023) SEMICON west, smart manufacturing: EcoTwin—an integrated solution
for sustainability in semiconductor manufacturing, San Fransisco, CA
86. Applied Materials inc. (2023) "EcoTwin™ Eco-efficiency software. [Online]. Avail-
able: https://www.appliedmaterials.com/us/en/semiconductor/solutions-and-software/ai-x/eco
twin.html
87. Applied Materials inc. (2023) Driving a net zero 2040 playbook powered by collaboration,
clean energy and innovation. [Online]. Available: https://www.appliedmaterials.com/us/en/cor
porate-responsibility/planet/net-zero.html
88. Lam Research inc. (2023) 2050 net zero strategy. [Online]. Available: https://www.lamres
earch.com/company/environmental-social-and-governance/2050-net-zero-strategy/
89. ASMI (2021) ASM international aims to achieve net zero by 2035. [Online]. Available: https://
www.asm.com/press-releases/2304062
90. NIST-IAC, (2023) Industrial advisory committee (IAC)—R&D gaps working group. [Online].
Available: https://www.nist.gov/system/files/documents/2023/02/08/Feb%207%20IAC%20M
eeting%20R%26D%20Gaps%20WG%20Materials%20Final.pdf
91. NIST (2023) NIST seeks nominations for the industrial advisory committee. [Online]. Avail-
able: https://www.nist.gov/news-events/news/2023/05/nist-seeks-nominations-industrial-adv
isory-committee
92. Moradian A, Neville E, Kelkar UM, Denome MR, Kothnur P, Ramanathan K, Shah K, Trejo
O, Meirovich S (2021) United States Patent US20220334569A1
93. Moradian A, Kelkar UM, Neville E, Trejo O, Meirovich S, Shah KB, Kher SS (2021) Uinted
States Patent US20230185268A1
94. Annepannavar A, Gopalakrishnan D (2021) Digital Twins for sustainable production:
modelling and simulation of a production system towards a Digital Twin. Chalmers University
of Technology, Gothenburg
95. Li L, Mao C, Sun H, Yuan Y, Lei B (2020) Digital Twin driven green performance evaluation
methodology of intelligent manufacturing: hybrid model based on fuzzy rough-sets AHP, II.
Complex Spec Issue: Complex Econ Bus 4:2020
96. Burkacky O, Göke S, Nikolka M, Patel M, Spiller P (2022) Sustainability in semi-
conductor operations: toward net-zero production, McKinsey & Company. [Online].
Available: https://www.mckinsey.com/industries/semiconductors/our-insights/sustainability-
in-semiconductor-operations-toward-net-zero-production
Digital Twins for Robot Systems
in Manufacturing

Ali Ahmad Malik , Guodong Shao , and Jane Tarakhovsky

Abstract The increasing need for industrial automation is driving the adoption
of robotics, with new developments such as human–robot collaboration through
autonomous mobile robots and collaborative robotic arms. While automation
improves product quality and working conditions and lowers manufacturing costs,
it can also limit manufacturing adaptability. Therefore, when integrating robots in
manufacturing, there is a pressing need to simplify the methods to develop, install,
reconfigure, and operate robot systems to achieve greater adaptability. This is where
the concept of Digital Twin, which replicates the behavior of a complex system in
a virtual environment for analysis and optimization, comes into play. In robotics,
the Digital Twin technology is expected to address the challenges associated with
designing, testing, commissioning, and reconfigurations. This requires a Digital Twin
to accurately represent various dimensions of a robotic system under production
variables. This study characterizes the components of a robot system that need to be
modeled in a Digital Twin to create a trustworthy virtual replica of a physical robot
system. A Digital Twin of this kind can be utilized throughout the lifecycle of the
physical robot installation across various use cases.

Keywords Robotics · Digital Twin · Flexible automation · Manufacturing

systems · Simulation · System design

A. A. Malik (B) · J. Tarakhovsky

Department of Industrial and Systems Engineering, Oakland University, Rochester, MI 48309,
USA
e-mail: [email protected]
J. Tarakhovsky
e-mail: [email protected]
G. Shao
Engineering Laboratory, National Institute of Standards and Technology, Gaithersburg,
MD 20899, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 307
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_13
308 A. A. Malik et al.

1 Introduction

Modern manufacturing systems are complex, with facilities evolving into large
networks of data-connected mechatronic components [1]. The complexity of these
systems stems from the quantity of information that spans various lifecycle phases,
including design, development, commissioning, operations, and end-of-life [2, 3].
The elevated complexity and interconnectivity make the lifecycle management
of modern-day manufacturing systems more challenging. Conversely, demand is
increasing for manufacturing systems to possess resilience and adaptability [4].
Addressing these challenges in complex scenarios requires enabling smart manu-
facturing through digitalization, data connectivity, and the integration of machine
learning [5]. Smart manufacturing, besides resilience, can bring cost reduction,
enhance workers’ well-being, and result in a better return on investment (ROI).
Industry 4.0, or the fourth industrial revolution, is the net sociotechnological
impact of infusing emerging technologies such as additive manufacturing, machine
learning, robotics, simulations, and the Internet of Things in products and their manu-
facturing systems. Advanced robotic automation stands out as one of the enablers
for Industry 4.0 [6]. Modern installations strive to make robotic automation more
flexible, adaptable, safe, and cost-effective than traditional robotics implementa-
tions. However, flexible approaches are lacking in developing plug-and-play hard-
ware, programming the robots, control program generation, task scheduling, layout
planning, safety assessment, and alignment with production plans.
Advancements in virtualization, sensing technologies, and computing power facil-
itate the realization of Digital Twins (DTs), which enable the testing and validation
throughout the design, development, and control phases of a complex system in a
virtual space. Different scientific domains increasingly recognize the potential value
of DTs for managing complexity in areas such as manufacturing, transportation,
aircraft, and space missions [7]. Manufacturing customization and reconfigurability
are vital domains to manage through DTs [7].
Computer models provide a means to shorten the time needed to design, redesign,
and deploy robot systems. Computer-based virtual models of physical systems can be
beneficial for testing and validating the production before implementation [8]. While
this method is consistent with traditional virtual modeling, the emerging “lifecycle”
approach and real-time communication between physical and virtual systems are
pivotal concepts of DTs [9].
Many studies have documented the potential advantages and relevance of
employing DTs for robot systems [10–12]. It has also been observed that devel-
oping a trustworthy virtual replica of a robot system is time-consuming and
demands advanced engineering skills and investment in different engineering soft-
ware. Creating and deploying a DT of a robot system should be structured, simplified,
and standardized to realize the needed ROI. It requires identifying the components of
a robot system that are relevant to the purpose of its DT. Moreover, the flexibility of
the DT itself is critical to ensure that the DT can effortlessly be adapted to evolving
circumstances.
Digital Twins for Robot Systems in Manufacturing 309

This chapter presents the importance of DTs in robotic installations within manu-
facturing systems. The components outlined in a DT of a robot system can assist
researchers and practitioners in developing cost-effective, modular, and flexible DTs,
thereby improving the resilience of robot installations. This is an essential step toward
achieving adaptable manufacturing systems.
The key contributions of this chapter are to:
1. Present the components of a DT of a robot system for flexibility
2. Examine the lifecycle phases of a robot system that a DT can support
3. Apply DTs in robot systems in manufacturing settings
4. Present use cases that demonstrate the utilization of DTs in robot systems.

2 Challenges and Opportunities in Contemporary

Manufacturing

The continuous drive to shorten product life cycles emerges as a significant transfor-
mation in the contemporary business landscape [13]. Emerging sociotechnological
trends require shorter product development and launch timelines. In this setting,
manufacturing companies leverage emerging hardware and software technologies
and their potential opportunities [14, 15].
Aside from the rapid pace of changes, manufacturers face a shortage of skilled
workers. The recent global exposure to the COVID-19 pandemic also displayed
widespread disruptions in supply chains [16] partly because of a shortage of workers
due to social distancing measures. Research studies identified that future factories
could better address such challenges by adopting modular, flexible, and human-
friendly automation solutions [17].
A way to develop human-friendly automation solutions is through flexible, collab-
orative robots. Technologies that facilitate the swift validation of new manufacturing
strategies are also needed. Therefore, future manufacturing systems must not only
be repurposable but also be designed, developed, commissioned, and reconfigured
at a significantly faster pace [18].
DTs can be utilized to address the resilience requirements within a manufac-
turing system. For example, DTs can help reduce the time required to validate
new manufacturing strategies, generate automation programs, and provide main-
tenance support. Additionally, DTs can harness real-time and historical data to offer
insights for process optimization. Such assistance can potentially enhance the level
of resilience that a manufacturing system can provide.
310 A. A. Malik et al.

3 Robotic Automation in Manufacturing

Automation describes assigning physical and cognitive tasks to machines and soft-
ware to boost production and decrease human effort [19] In manufacturing, adopting
automation brings advantages such as enhancing workplace safety, efficiency, quality,
and cost-effectiveness [20, 21]. However, this often comes at the cost of reduced
production flexibility. At the heart of industrial automation lies industrial robots.
The subsequent sections elaborate on the diverse types of robots employed in manu-
facturing facilities. Figure 1 shows various industrial robot types, including spher-
ical, SCARA (Selective Compliance Assembly Robot Arm), delta, Cartesian, and
humanoid robots [22]. The robot selection for a specific task is based on the nature of
the tasks to be automated, available space, financial constraints, and process-related
considerations. While these robots enhance manufacturing efficiency, their applica-
tions are limited in certain operations, such as assembly, which only constitutes 7.3%
of robotic use [23].

3.1 Traditional Industrial Robots

Robots are the predominant force driving the industrial automation of physical tasks
[22, 24]. These robots, characterized by fixed positioning, operation within enclo-
sures, and time-consuming reconfiguration processes [25], fall into the category of
fixed automation solutions. They can help achieve high production volumes but
must strictly be separated from human interaction. They also demonstrate limited
flexibility [26]. Industrial robots have proven successful in various sectors, including
automotive, medicine, food, and electronics manufacturing [27]. The primary reason
for the unsuitability of robots in assembly is human safety and the challenges of their
reconfiguration [28].

Fig. 1 Various types of industrial robots

Digital Twins for Robot Systems in Manufacturing 311

Fig. 2 A human and robot coexisting in manufacturing [36]

3.2 Collaborative Robots

Modern industrial robots are lighter, portable, easy to program, and safe. This change
can meet the need for flexibility in terms of mobility, capability, and capacity [29].
These robots, designed for collaboration and coexistence with humans (Fig. 2), are
commonly known as “cobots” or collaborative robots [30]. A collaborative robot
can be defined as a mechanical device intended for direct physical interaction with
humans, a concept first introduced by Colgate [31] and further developed by Kruger
[32]. Cobots allow humans and robots to work together to harness the strengths
of humans and machines. This concept, often called lean automation, exists at the
convergence of human flexibility and machine efficiency [33].
Literature showcases diverse applications of cobots, spanning pick-and-place
operations, assembly tasks, welding, inspection processes, and packing [34]. More-
over, cobots have been explored as a viable solution for rapidly repurposing factories
in emergencies [17]. The predominant use of cobots has been in manufacturing small
components such as those assembled into electronics, appliances, and electronic actu-
ators [28, 35]. With the advancement of sensing and safety technologies, cobots are
being considered to automate large and heavy components.

3.3 Autonomous Mobile Robots

Autonomous mobile robots (AMRs) represent a distinctive category within collabo-

rative robots [37]. They have proven highly effective in material handling applications
312 A. A. Malik et al.

in manufacturing settings [40]. Their adaptability and versatility make them well-
suited for tasks requiring interactions with humans. Beyond manufacturing, AMRs
have found applications in warehouses, military operations, healthcare, search and
rescue missions, security, and home environments [38]. This versatility underscores
the potential of mobile robots to automate operations in diverse fields.
Different robot types have standard features such as mechanical multijointed
reprogrammable actuators, end-of-arm tooling, machine vision, positioning tech-
nology, and control programs. Adaptability is recognized as necessary for most
modern-day robots. The following section presents a typical physical architecture of
a robot installation.

4 Architecture of a Robot System in Manufacturing

Robotic systems in manufacturing settings are available in various designs, layouts,

and configurations, influenced by specific use cases. A typical robot cell comprises
multiple hardware and software components (Fig. 3). Articulated robot arms, with
one or more reprogrammable mechanical joints, represent most robot installations
[39]. One or more tools (end effectors) are attached to a robot manipulator’s tool post
to perform various functions. A robot controller oversees the operations of the robot
system, connecting all external hardware through the input/output (I/O) interfaces
of the controller. Various sensors are embedded in the robot body and integrated
externally to monitor its performance and respond to emerging situations.
Robotic arms designed for human–robot collaboration (HRC) typically have
power and force-limiting bodies, speed and separation monitoring, hand guiding, and
emergency stop as stipulated in the ISO15066 safety standard for HRC [21]. These
attributes ensure safe interaction between robotic arms and humans [40]. AMRs can

Fig. 3 Typical components of a robot system in a manufacturing setting

Digital Twins for Robot Systems in Manufacturing 313

also facilitate the mobility of a robotic arm within the robot system. Furthermore,
collaborative robots must comply with the ISO19649 standard [41], which estab-
lishes terms related to mobile robots operating on solid surfaces and engaging in
industrial robot applications. To effectively employ the concept of DTs, it is vital to
model most, if not all, of these features of robot systems in their digital models.

5 DTs for Robots in Manufacturing

A DT is a virtual representation of the components and dynamics of an observable

physical system [44]. DTs can mirror real-time operating conditions and predict
the future behavior of a physical system [45]. The core concept of a DT involves
creating a digital model of a physical system and linking each component of the
digital model to its corresponding physical assets. In return, the virtual model must
act as a front-runner of the physical system to predict or estimate its future behavior.
The present understanding of the DT concept originates from the idea of a “Con-
ceptual Ideal for PLM” (Product Lifecycle Management) [42]. It proposes that every
system is a subset of two other systems: the physical system in the physical world
and a virtual system existing in virtual space, containing all necessary information
about the physical system. The bidirectional relationship between the physical and
digital systems can enhance product design, manufacturing, and service throughout
the system’s life cycle [43].
The methods of using an informational virtual model to represent the complexity
of a physical system have evolved. In earlier times, the virtual model existed as a
mental image [1], limited in its capacity to address questions about the system’s
performance. In the mid-twentieth century, creating virtual models became possible,
starting with 2D CAD (computer-aided design) objects and advancing to 3D models
and dynamic simulations. These virtual models are typically developed early in the
system’s lifecycle, i.e., during design. These models often become useless when the
system transitions to the operation phase.
The linkage of digital models to their physical counterparts and the integrated
intelligence throughout their lifecycle are now achievable (Fig. 4). This enables
them to understand operational behavior and assist in addressing day-to-day produc-
tion constraints. In this context, DTs can be categorized into DT prototypes and DT
instances [44]. A DT prototype is used to refine system design, presenting optimal
static and dynamic information to achieve desired outcomes. Meanwhile, a DT
instance integrates monitoring, service, sensing, and behavioral information about
the physical twin during its operations. DT instances exhibit predictive and inter-
rogative behaviors, which prove beneficial during the operational and maintenance
phases.
314 A. A. Malik et al.

Fig. 4 Scope of DTs in manufacturing systems

6 Lifecycle Phases of Robot Systems

A robotic system undergoes a comprehensive life cycle, commencing with its design
and concluding at its end-of-life stage. Correspondingly, its DT follows a parallel life
cycle, adapting to various scenarios and system evolution throughout the lifecycle
(Fig. 5). The subsequent section describes the functions of a DT across multiple
stages in a robotic system’s life cycle.

Fig. 5 Concept of a Digital Twin system in human–robot collaboration

Digital Twins for Robot Systems in Manufacturing 315

6.1 Design of the Robot System

In developing a new robot system, it is customary to construct virtual models before

the actual physical counterpart is built. This virtual representation, which can also be
referred to as a DT prototype, is conceived to conceptualize and finalize the system’s
appearance, specifications, the selection of off-the-shelf components, and the overall
layout. Despite the absence of the corresponding physical counterpart during the
design phase, the DT represents the intended physical twin. It enables the exploration
of various what-if scenarios, facilitating swift, secure, and reliable design outcomes.
The choice of robot manipulators, workstation design, layout, fixtures, and financial
assessments are critical questions that must be addressed at the design stage.

6.2 Commissioning the Robot System

The results derived from the design phase provide information for developing the
components of the physical system. The developed system then moves to the commis-
sioning stage. In the case of a robot system, this stage may entail the creation of work-
stations, fixtures, feeding devices, and other hardware elements. The Bill of Materials
(BOM) and Bill of Processes (BOP) can be generated, guiding the development of
the physical system. Throughout this phase, the connection between the physical
systems and their DTs can be established by linking the DT to an actual controller or
programmable logic controller (PLC) to identify potential errors. This methodology
is analogous to virtual commissioning (VC). VC, or hardware-in-the-loop simula-
tions, reduces development time by facilitating virtual testing and integration well
before actual commissioning. The physical robot can be live connected with its DT,
allowing it to execute tasks as designed in the DT.

6.3 Scheduled and Preventive Maintenance

Maintenance is an essential component of most production systems. Emerging tech-

nologies such as augmented reality (AR) or chatbots can be integrated with a DT to
optimize maintenance procedures, which can better assist maintenance personnel
with enhanced visualization tools for fault detection and training tasks. Virtual
reality (VR) is another visualization technology that can be integrated with the DT,
particularly for training.
Maintenance can benefit from the DT technology in ways such as:
. Real-time Monitoring: Data capturing operating parameters, energy consumption,
and system health.
. Predictive Maintenance: Use of machine learning within DT to predict potential
failures or maintenance needs based on performance trends.
316 A. A. Malik et al.

. Condition Monitoring: IoT sensors provide data on system health indicators,

which can be integrated with DT for continuous condition monitoring.
. Asset Tracking: The usage and lifecycle of robotics assets, such as operating hours,
replacement history, etc., can be tracked and help with proactive maintenance
scheduling.
. Remote Diagnostics and Troubleshooting: Identifying issues remotely can help
reduce downtime and improve overall efficiency.

6.4 Operations and Changeovers

The most compelling application of DT technology lies in its application throughout

the operational life of a robot system. Over time, a robot system may need
changeovers, safety assessments, production analyses, and modifications. Assessing
the overall equipment effectiveness (OEE) is another practical facet when evaluating
a production system for continuous performance optimization. A DT plays a pivotal
role in elevating the quality of these processes, thereby enhancing the performance
and reliability of the robot system. This helps justify the investment in the creation
and maintenance of its DT.
The DT developed during the design phase is extended to facilitate real-time
communication with the physical system during operation, enabling behavioral
analysis and performance optimization. At this stage, the system synchronizes the
real-world data with the DT, enabling automated assessment cycles. This cyber-
physical system integrates production planning and control databases to support
scheduling production orders and changeovers. The DT is valuable in simplifying
the reconfiguration or repurposing of the robot system in response to demand
fluctuations.

7 Components of a DT for a Robot System

This section presents the fundamental components or modules comprising a robotic

system’s DT, as illustrated in Fig. 6. Traditionally, various tools are required to
simulate each of these components. Connectivity protocols can enable near real-
time communication between these components, streamlining the development of
an accurate DT for a robotic system.

7.1 Static CAD Modeling

The initial step in constructing a virtual manufacturing system is to create a 3D

visualization using CAD software. There is a multitude of tools available for this
Digital Twins for Robot Systems in Manufacturing 317

Fig. 6 Components of DT of robot systems

purpose. This CAD data can be obtained directly from the equipment manufacturer,
often in standard exchangeable formats such as STEP (Standard for the Exchange
of Product Data) [45]. Robot manufacturers offer CAD models of their robots, and a
similar practice is followed by manufacturers of related equipment, such as grippers,
fixtures, feeders, and tables. Furthermore, many simulation tools feature a built-in
library of proprietary and generic factory resource CAD models.
A critical step in preparing the CAD data is creating an assembly file and consol-
idating the individual CAD models of various devices and equipment. The virtual
assembly model must represent the complete physical robot system being investi-
gated. Each component can be assigned material properties and visualization to aid
in subsequent analyses. This file can be exported to various exchangeable formats,
with STEP being the most common standard format.

7.2 Process Simulation

The CAD data can be imported into a continuous simulation environment. Creating a
dynamic simulation starts with defining the kinematics of each active resource within
the system. It involves specifying position and location constraints, joint types, joint
limits, and velocity limits. For example, a gripper may need to be defined for its
318 A. A. Malik et al.

motion kinematic joint types, limits, and action poses. The visualization/simulation
of a DT is achieved through three steps: (1) creating the simulation model of a
robot system along with its operation sequences, akin to a Gantt chart, (2) an event-
driven continuous simulation that runs for a pre-determined time and controlled by
an internal logic engine, and (3) the simulation is controlled through signals from a
virtual PLC and other emulators. This simulation becomes the primary component of
the DT for visualization, experimentation, and analysis. After the simulation, it can
perform analyses (e.g., collision detection, layout assessment, cycle time estimates)
and optimizations. Numerous proprietary tools are available to create this type of
simulation, while open-source engines can also be utilized.

7.3 Automation Program

PLCs serve as industrial computers for programming and monitoring industrial robot
systems. A critical step in commissioning a robot-based manufacturing system is
creating and validating the automation program. Usually, this program is created
later in the development stages. Developing, testing, and validating the automa-
tion program in a virtual space, along with process simulation, enhances the relia-
bility of the system’s performance. Each PLC has its programming tool, and open-
source program development tools are also available. To ensure interoperability,
PLC programs follow the IEC 61131 standard [46]. The IEC 61131-3, developed
by the International Electrotechnical Commission (IEC), sets the standard for PLCs’
syntax, semantics, and interoperability. The developed programs are downloaded
onto a virtual PLC and interfaced with the simulation.

7.4 Mechatronic Behavior

A robot system includes sensors, actuators, feeders, fixtures, and other mecha-
tronic elements (Fig. 3). Behavioral modeling of these devices enables an accurate
virtual model of the entire system. The Functional Mock-up Unit (FMU) is a tool-
independent, free standard crafted for dynamic model exchange and co-simulation.
FMUs define a container and an interface for sharing dynamic simulation models
through a combination of Extensible Markup Language (XML) files, binaries, and
C-code. Both commercial and open-source tools are accessible for simulating the
behavior of each device and interfacing it with the process simulation.
Digital Twins for Robot Systems in Manufacturing 319

Fig. 7 Production assessments through simulations

7.5 Production Variables

The evaluation of production variables based on historical statistical data is not always
required in the process modeling of a robot system. However, it is a component that
can be added to a DT of a robot system for production-related analysis. In this phase,
production-related parameters and throughput specifications are defined. Since these
analyses are done in a stochastic simulation, a different tool is often required to
perform stochastic modeling and interface it with the existing continuous process
simulation model. This model simulates cycle time, startup time, setup time, potential
failure scenarios, repair and maintenance requirements, shifts, worker allocation, and
other relevant factors with statistical probabilities. An example of such a simulation
is shown in Fig. 7, where six human–robot packaging stations are shown, and the
simulation presents the operational time and waiting time for each resource. Discrete
simulations can run thousands of trials based on probabilistic distributions derived
from historical data. This analysis enables the prediction of the throughput of a robot
system under the design variables.

7.6 Human–machine Interface

In industrial settings, human–machine interfaces (HMIs) enable workers to interact

with manufacturing systems or robots. This interaction involves conveying instruc-
tions such as start/stop commands, speed adjustments, and troubleshooting. Touch-
screen HMIs are commonly employed in industrial settings for this purpose. An
HMI is essential for facilitating continuous user interaction with the robot system in
most robot work cells. Depending on the process and system design, modeling and
integrating an HMI with the process simulation may be needed in a DT. A simulated
HMI can communicate with PLCs and the process simulation. Using a virtual HMI
320 A. A. Malik et al.

allows end users to interact with the robot system in a way that is similar to its actual
application. HMIs are designed following the ISA 101 standard, and validating them
may require a DT model for thorough validation. A virtual HMI may be accessed on
a computer screen or a handheld computing device communicating with the process
simulation. The HMI developed during this step is downloadable to real HMIs for
practical field applications.

7.7 Data Communication and Management

In robot system DTs, real-world data are integrated into the simulation. This integra-
tion enables continuous assessments under variable conditions, enhancing the accu-
racy and responsiveness of the DT. System performance data, including logs and
alerts, can be stored in a data repository for ongoing evaluations. Various sensors,
tailored to specific requirements, can be employed in robotic systems to log parame-
ters such as robot joint positions, machine vision data, collision events, task comple-
tion status, safety breaches, and cycle time. As exemplified by [1], a robot assembly
system is connected to a cloud data repository through an internet-based router for
data logging. It is an HRC assembly cell using a UR-5 robot. The performance-
related logs from the assembly cell are stored and used in the simulation for design
optimization (Fig. 8). These performance factors include the idle time for both the
operator and the robot, human operator safety (collision occurrences), and completed
cycle counts. The recorded data is then used for layout and robot path optimization
in the simulation.
Data management organizes data into meaningful information, focusing on
creating “Golden datasets” that undergo cleaning, transformation, validation, integra-
tion, and standardization. The data management lifecycle encompasses key stages:

Fig. 8 Communication of data logs from robot system [47]

Digital Twins for Robot Systems in Manufacturing 321

data collection, transmission, storage, processing, fusion, and visualization. During

data collection, hardware, software, and network resources are integrated. Data trans-
mission involves various wireless and wired technologies. Data storage, required
for processing and analysis, utilizes technologies such as cloud storage, NewSQL
databases, NoSQL databases, and distributed file storage (DFS). After processing
data, data visualization can be made available in formats required by AR or VR [48].

8 Applications of DTs in Robots

Deploying and utilizing DTs alongside robot systems in manufacturing facilities

offers a range of benefits. This section presents various advantages of robot systems
contributing to robot installations’ flexibility, reliability, and safety.

8.1 Safety Assessment and Validation

Manufacturing-related robot installations comply with safety standards, notably ISO

10218 [49] and ISO 13849 [50], which outline safety requirements for industrial
robots. Moreover, collaborative robots align with safety specifications outlined in
ISO 15066 [51], which is dedicated to collaborative robotic devices. Various safety
measures, including emergency stops, human movement monitoring, and active colli-
sion avoidance, are integrated into robot cells to safeguard coexisting humans. The
assessment and validation of robot systems for compliance with safety requirements
are essential.
The DT model should accurately replicate the robotic system’s physical charac-
teristics, behaviors, and interactions (movements, sensors, and environment) to be
valid. Under ISO 15066, a safety risk assessment is required following any physical
alterations to the system. The risk assessment process can be streamlined using DTs
within a controlled and risk-free environment. Risk identification can be conducted
in the DT, maintaining a live connection with the physical robot. This approach
identifies potential collisions in the virtual environment before the physical system
is populated with hardware resources.
Evaluating potential accidents or injuries is essential in the context of HRC.
Dynamic simulations within DT can evaluate safety performance under different
operating scenarios (emergency, failure modes, and normal operations). Collision
detection algorithms within the DT can identify potential failures and develop
strategies for avoidance or mitigation to enhance safety.
Continuous monitoring using collision-related data logs can enable the assessment
of the frequency of collisions between humans and robots over time, facilitating the
optimization of robot paths to avoid such incidents. Fault tolerance mechanisms
and redundancy in a DT can enhance safety and reliability. The DT can also offer
training and awareness programs for operators, maintenance personnel, and other
322 A. A. Malik et al.

stakeholders. The DT models can also simulate the safety training scenarios within
a virtual environment to enhance learning and preparedness.

8.2 Reduced Development Time

Studies have found that robotic automation projects frequently exceed the initially
projected timeline [52]. This duration can be further extended, particularly with
more complex tasks such as assembly and battery pack manufacturing. Challenges
that emerge unexpectedly in the planning stage, impacting the project timeline,
include issues related to process balancing, task scheduling, feeding methods,
fixtures required, the necessity for multiple grippers, and safety complications.
This prolonged integration timeframe is attributed to the nature of robot operations.
Robots must coordinate their movements with other hardware (machine tools, equip-
ment, end-of-arm tooling), peripherals (vision systems, force sensors), and humans,
contributing to increased integration and operational complexity.
Various methods and frameworks have been documented for developing robot
systems [53–55]. However, there is a growing need for novel approaches that focus
on minimizing the time and effort required for integration, validation, and reconfig-
uration. DT prototypes can offer a high-fidelity and trustworthy digital model of the
real system to reduce the chances of any errors that may arise at a later stage.

8.3 Robot Programming

Robot programming involves defining the paths, actions, and logical procedures for
robots to perform the assigned tasks. Industrial robotic applications often require
significant expertise and effort. Despite the promise of more accessible program-
ming in the latest robots, manually programming complex robot paths remains
time-consuming. With the continuous desire for customization and changing market
dynamics, robot programming is not a one-time activity in their operational life. The
needed flexibility and adaptability require easy ways of programming the robots.
To tackle this challenge, a DT can facilitate the intuitive development of robot
programs offline within a graphical environment (Fig. 9). The robot program can
be seamlessly downloaded from the DT directly to the connected robot by virtually
testing the desired operation, complete with defined robot trajectories and logic. This
approach streamlines the programming process, leading to a significant enhancement
in efficiency during robot deployment.
Digital Twins for Robot Systems in Manufacturing 323

Fig. 9 Generating robot program and controlling robot arm through its DT [47]

8.4 Simplified Reconfigurations

The demand for variety in manufacturing is rising, and assembly is recognized as a

prime area with significant potential for introducing product diversity [56]. As we
navigate the transition phase of the fourth industrial revolution, the notion of batch
size one, focusing on personalized manufacturing, is increasingly gaining relevance.
However, a challenge is that current robots often lack the flexibility necessary for
rapid reconfiguration.
To realize the vision of batch size one, manufacturers require robots with features,
hardware, and software solutions that facilitate effortless adaptation to changes and
variations in product design. Swift reconfiguration of robots is indispensable for
staying in sync with the dynamic demands of modern manufacturing, where rapid
shifts in product specifications and design variations are commonplace. DT models
emerge as a valuable tool in expediting these reconfigurations by automatically eval-
uating the automation potential for specific tasks. This includes generating a process
sequence, automation programs, robot codes, and worker training opportunities. DT
models contribute to streamlining the adaptation process, ensuring efficiency and
agility in response to evolving manufacturing requirements.
324 A. A. Malik et al.

8.5 Layout Planning and Optimization

In a robot installation, achieving an optimal arrangement of robots, fixtures, assembly

parts, and associated hardware is critical to minimize footprints and shorten cycle
times. In the case of collaborative robots, additional safety layers need attention. To
enhance the design of the workstation layout, the following experiments can play a
pivotal role:
Collision analysis: Identifying and mitigating collisions become paramount in the
robotic process in scenarios involving confined spaces with multiple robots and
associated hardware. Conducting a thorough collision analysis serves to identify
potential collisions, allowing for the optimization of robot trajectories and equipment
placement. Given the frequent changeovers and reconfigurations inherent in such
environments, this collision analysis may need to be a routine activity. A DT emerges
as a valuable tool for generating safe robot paths that circumvent potential collisions
(see Fig. 10). By using DTs, robot paths can be proactively optimized. The result is a
more efficient and collision-free operation, even in dynamically changing scenarios.
Reach test: The reach test is instrumental in determining whether a robot can access
all desired locations within its workspace. This test aids in defining the most optimal
locations for placing robots and relevant production equipment. Specific location
points for robots are established, evaluating whether the robot can safely reach all
desired locations. A grasp envelope is created for a human reach test to demonstrate
the human arms’ reachability without adopting unsafe body postures.
Placement test: The placement test seeks to identify optimal locations for robots,
humans, and production equipment. The goal is to minimize cycle times, prevent

Fig. 10 Layout planning of a robot cell in simulation

Digital Twins for Robot Systems in Manufacturing 325

collisions, and ensure a safe working environment. The test can define a range of
points from which a robot can reach a selected set of locations.
Other considerations may include designing the layout with flexibility and scal-
ability to accommodate future changes and expansions. Modular design approaches
and flexible workspace configurations enable easier reconfiguration and adaptation to
evolving production needs. Another important consideration is ensuring the seamless
integration of robotic installations with existing manufacturing systems, including
conveyor systems and automated guided vehicles. Coordinating the layout design
with other production processes (inventory management systems) is essential to
optimize workflow integration and synchronization.

8.6 Process Visualization

A comprehensive simulation enables high-accuracy visualization of the detailed

process. Even without making any computer-based analysis of collisions, robot path,
or layout visualization, visualization allows visual assessment of the process. Poten-
tial errors are identified, and the required modifications are determined. As shown in
Fig. 11, a simulation of a robotic process presents the assembly of medical ventilators.

Fig. 11 Sketch shows the assembly process sequence in a robotic assembly process
326 A. A. Malik et al.

8.7 Control Program Generation

Robot systems with multiple robotic arms, sensors, actuators, feeders, and hard-
ware devices typically operate based on a logical control program. Traditionally,
the development and verification of this logical program are carried out during the
commissioning phase. The validation and error-proofing procedures can be time-
consuming and may uncover errors that could set the project back to earlier stages.
However, leveraging a DT allows for a more streamlined approach. The control
program can be generated directly from the simulation model and verified several
times during the system development. This generated program is then downloaded to
a PLC. The advantage of this approach is that, with each change in the system design,
a new control program can be rapidly created, validated within the DT simulation,
and then efficiently downloaded to the PLC. This dynamic adaptability enhances the
efficiency and agility of the robot system, reducing the time and effort traditionally
associated with the commissioning process.

8.8 Assessment of Return on Investment

ROI analysis is crucial for decision-making when implementing a robot system to

automate a specific process. Without a detailed visualization, analysis, and under-
standing of operational behavior, accurately assessing the financial gains from
investing in a robotic system becomes challenging. A comprehensive simulation-
based DT offers an in-depth perspective on various phases of a robot system’s life-
cycle. It provides a more accurate and thorough assessment of ROI throughout its
entire operational lifespan.

9 Practical Use Cases

This section presents two use cases from industrial practices that exemplify the
practical application of DTs in robot systems.

9.1 DT-Based Development of a Robot Assistant in Wind

Turbine Manufacturing

This case exemplifies the creation and utilization of a DT for a collaborative mobile
robot aiming to automate assembly tasks in the manufacturing of wind turbines.
Applying a mobile robot assistant for hybrid automation of wind turbine manufac-
turing is expected to give benefits such as reduced production costs, enhanced product
Digital Twins for Robot Systems in Manufacturing 327

Fig. 12 DT-supported deployment of a robot assistant in wind turbine manufacturing

quality, and improved working conditions. The DT technology was employed for the
commissioning and reconfiguring the mobile robot, contributing to expedited design
and validation.
The growing demand for higher rating capacity in wind turbines has driven the
production of larger generators, extended blades, and taller towers. The assembly
of these turbines entails the manipulation of large-sized components. Traditional
automation is impractical due to frequent design alterations and diverse tasks,
rendering it labor-intensive. However, human capabilities fall short in managing
the assembly of these substantial components, resulting in extended lead times and
increased overall costs for sustainable energy. To address these automation chal-
lenges, the case presents a solution consisting of a mobile robot equipped with
a robotic arm that uses laser beams to assist operators in precisely positioning
assembly components. This approach streamlined the assembly process and reduced
the assembly time.
The simulated robot was connected to the physical robot in real time using an
Ethernet connection (see Fig. 12). The robot’s IP address was utilized in the simula-
tion to establish this connection. A post-processor for Doosan Robotics was used. A
robot post-processor defines how robot programs are generated for a specific robot
controller.
During the design phase, a comprehensive workspace with the robot system was
modeled in simulation-based DT. The DT model served as a design validation tool for
each reconfiguration. During the development phase, the DT model was a reference
for programming the robot. Assembly locations were extracted from the DT model
and encoded in the robot program. Vision tests and safety risk assessments were
performed using the DT. The DT was operational with the robot system to verify the
changes and robot programming.

9.2 DT of a Human-Robot Collaborative Assembly Cell

This section illustrates the application of a DT in developing and operating a collabo-

rative assembly cell (Fig. 13), emphasizing its significance in designing, developing,
and operating human–robot production systems. In the design phase, the DT was
328 A. A. Malik et al.

Fig. 13 Humans and robots working collaboratively [47]

used to create and select the elements of the assembly system according to produc-
tion requirements and its integration with the overall system. Dynamic simulation
facilitated a quantitative assessment and a business value proposition for the proposed
solution. The Task Simulation Builder in TPS was used to model human tasks for
ergonomic assessments.
During the operational phase, using data loggers, the DT transitioned from manual
data syncing to automated and real-time data syncing, enhancing its utility for error
identifications. The DT proved valuable in dynamic task distribution based on task
complexity rating and event-driven simulation. Other benefits included intuitive robot
programming to reduce manual efforts, human safety assessments, the generation of
data logs for critical actions through sensor integration, and the incorporation of
artificial intelligence to enable the system to self-learn and make decisions based on
past experiences.

10 Future Research Directions

Interoperability stands out as a critical challenge in the current state of creating DTs
for robot systems. The various aspects of a robot system (geometry, kinematics,
robot program, automation logic, etc.) can be modeled using different tools (e.g.,
Robot Operating System (ROS) [57], OpenPLC [58], Unreal Engine [59], etc.).
However, exchanging data between these tools is often complicated and sometimes
impossible. Additionally, commercial tools from one vendor may not exchange data
with those from another. Consequently, there’s a pressing need for standardized,
exchangeable data formats for DT devices and assets, particularly their connectivity
with open-source tools. Initial standards developments such as the IEC 63278—
Asset Administration Shell [60] have been made. Still, they are at an early stage
Digital Twins for Robot Systems in Manufacturing 329

of development, and their widespread adoption needs to address their suitability for
various scenarios.
Creating and operating a DT necessitates investment in software tools, human
resources, data processing ability, and communication technologies. However, it is
essential to recognize that not every robot installation requires a DT. Fundamentally,
a DT benefits robotic systems as it facilitates flexibility, reconfigurability, seamless
deployment, predictive maintenance, and safe human–robot interaction, among other
benefits. Assessment methods must be available to identify the need and value of a
DT in a given context and asses the ROI of creating and using a DT.
In manufacturing systems, hardware modularization involves incorporating
parallel modules for capacity or capability adjustments. Combining flexible automa-
tion with human skills, this modular approach facilitates reconfigurability and agility.
Achieving a high degree of customization involves adding, replacing, or eliminating
modules. A similar modular approach is needed in DTs. Given that the current
approach to developing DTs is time-consuming, the creation of libraries of modular
DTs of assets and information blocks can pave the way for the formation of reusable
and exchangeable DTs.
Flexible robot installations face the challenge of effective and dynamic task
scheduling. Task scheduling for robots involves the creation of an optimal schedule
within a system. This schedule specifies which robot is assigned to each task and
how the tasks will be processed. Dynamic task allocation goes a step further, encom-
passing the planning of automation processes and allowing robots to adapt during
operations. To address this challenge, a comprehensive approach that integrates
event-based logical simulation, probabilistic analysis, and statistical data analysis
into a DT model is needed. This integration can enhance the DT’s capability to
facilitate dynamic and efficient task scheduling, ensuring adaptability to changing
operational needs.

11 Conclusion

A manufacturing system can be characterized as a network of subsystems, including

equipment, machines, humans, and robots, working together to transform raw mate-
rials into finished products. Robots have become integral components of modern
manufacturing facilities and are increasingly gaining traction. However, designing,
developing, deploying, and reconfiguring robot cells is time-consuming, spanning
weeks to months. The extensive involvement of a larger workforce in the project
contributes to higher overall system costs and results in a prolonged ROI. This
extended timeframe is influenced by various factors, such as sequential develop-
ment processes, complex programming techniques, tool development, and the need
to adhere to safety standards.
This chapter outlines the fundamental elements or modules constituting a robotic
system’s DT. The research arena has documented the advantages of employing DTs
for robot systems. However, realizing the full potential of DTs for robot work cells
330 A. A. Malik et al.

requires the creation of a reliable and comprehensive digital representation that accu-
rately models the elements and dynamics of the observed robot system. Flexibility
remains critical in ensuring the DT can seamlessly adapt to changing circumstances.
This chapter underscores the significance of DTs in manufacturing robotic cells and
outlines the essential criteria for developing these DTs.
The interoperability challenge arises in creating DTs for robot systems, as
exchanging data between different tools proves difficult. Standardized and exchange-
able data formats are essential, emphasizing compatibility with open-source tools.
While investing in DT tools and resources is crucial, it’s important to note that
not every robot system requires a DT. Modularization is critical in manufacturing
and DT development, offering flexibility and reconfigurability. Flexible robot instal-
lations encounter challenges in dynamic task scheduling, which can be addressed
by integrating logical simulation and statistical analysis into DT models, ensuring
adaptability to changing operational needs.

Disclaimer This research was conducted through the support of a NIST cooperative agreement
[60NANB23D234]. Specific commercial products and systems are identified in this paper to facil-
itate understanding. Such identification does not imply that these software systems are necessarily
the best available for the purpose. No approval or endorsement of any commercial product by NIST
is intended or implied.

References

1. Malik AA, Brem A (2021) Digital twins for collaborative robots: a case study in human-
robot interaction. Robot Comput Integr Manuf 68:102092. https://doi.org/10.1016/j.rcim.2020.
102092
2. Rios J, Bernard A, Bouras A, Foufou S (2017) Product lifecycle management and the industry
of the future. In: 14th IFIP WG 5.1 international conference. Springer, Seville, Spain
3. Morgan J, Halton M, Qiao Y, Breslin JG (2021) Industry 4.0 smart reconfigurable manufacturing
machines. J Manuf Syst 59:481–506. https://doi.org/10.1016/j.jmsy.2021.03.001
4. Yazdani MA, Khezri A, Benyoucef L (2022) Process and production planning for sustainable
reconfigurable manufacturing systems (SRMSs): multi-objective exact and heuristic-based
approaches. Int J Adv Manuf Technol 119:4519–4540. https://doi.org/10.1007/s00170-021-
08409-0
5. Shao G, Helu M (2020) Framework for a digital twin in manufacturing: scope and requirements.
Manuf Lett 24:105–107. https://doi.org/10.1016/j.mfglet.2020.04.004
6. Advanced Manufacturing (2022) National strategy for advanced manufacturing
7. Malik AA (2023) Simulation based high fidelity digital twins of manufacturing systems: an
application model and industrial use case. In: Winter simulation conference 2023
8. Masood T, Weston R (2011) An integrated modelling approach in support of next generation
reconfigurable manufacturing systems. Int J Comput Aided Eng Technol 3:372–398
9. Bilberg A, Malik AA (2019) Digital twin driven human–robot collaborative assembly. CIRP
Ann 68:499–502. https://doi.org/10.1016/j.cirp.2019.04.011
10. Maruyama T, Ueshiba T, Tada M, Toda H, Endo Y, Domae Y (2021) Digital twin-driven human
robot collaboration using a digital human 1–46. https://doi.org/10.3390/s21248266
11. Liu S, Wang XV, Wang L (2022) Digital twin-enabled advance execution for human-robot
collaborative assembly. CIRP Ann 71:25–28. https://doi.org/10.1016/j.cirp.2022.03.024
Digital Twins for Robot Systems in Manufacturing 331

12. Yao B, Xu W, Shen T, Ye X, Tian S (2023) Digital twin-based multi-level task rescheduling
for robotic assembly line. Sci Rep 13:1–20. https://doi.org/10.1038/s41598-023-28630-z
13. Masood T, Kern M, John Clarkson P (2021) Characteristics of changeable systems across value
chains. Int J Prod Res 59:1626–1648. https://doi.org/10.1080/00207543.2020.1791997
14. Giones F, Brem A (2017) From toys to tools: the co-evolution of technological and
entrepreneurial developments in the drone industry. Bus Horiz 60:875–884
15. Stadnicka D, Antonelli D (2019) Human-robot collaborative work cell implementation through
lean thinking. Int J Comput Integr Manuf 32:580–595. https://doi.org/10.1080/0951192X.2019.
1599437
16. Malik AA, Masood T, Kousar R (2020) Repurposing factories with robotics in the face of
COVID-19. Sci Robot 5, (2020) https://doi.org/10.1126/scirobotics.abc2782
17. Malik AA, Masood T, Kousar R (2021) Reconfiguring and ramping-up ventilator production
in the face of COVID-19: can robots help? J Manuf Syst 60:864–875. https://doi.org/10.1016/
j.jmsy.2020.09.008
18. Anumbe N, Saidy C, Harik R (2022) A primer on the factories of the future
19. Nof SY (2009) Springer handbook of automation. Springer, Berlin
20. Boy GA (2018) From automation to interaction design. Handb Hum-Mach Interact 431–443.
https://doi.org/10.1201/9781315557380-22
21. Malik AA, Masood T, Management E, Brem A (2024) Intelligent humanoid robots in manu-
facturing: addressing the worker shortage and skill gaps in assembly cells. https://doi.org/10.
1145/3610978.3640765
22. Kurfess TR (2004) Robotics and automation handbook
23. Shneier MO, Messina ER, Schlenoff CI, Proctor FM, Kramer TR, Falco JA (2015) Measuring
and representing the performance of manufacturing assembly robots
24. Andersson SKL, Granlund A, Hedelind M, Bruch J (2020) Exploring the capabilities of indus-
trial collaborative robot applications. Adv Transdisciplinary Eng 13:109–118. https://doi.org/
10.3233/ATDE200148
25. Pedersen MR, Nalpantidis L, Andersen RS, Schou C, Bogh S, Krüger V, Madsen O (2016)
Robot skills for manufacturing: From concept to industrial deployment. Robot Comput Integr
Manuf 37:282–291
26. Ji S, Lee S, Yoo S, Suh I, Kwon I, Park FC, Lee S, Kim H (2021) Learning-based automation of
robotic assembly for smart manufacturing. Proc IEEE 109:423–440. https://doi.org/10.1109/
JPROC.2021.3063154
27. Buerkle A, Eaton W, Al-Yacoub A, Zimmer M, Kinnell P, Henshaw M, Coombes M, Chen
WH, Lohse N (2023) Towards industrial robots as a service (IRaaS): flexibility, usability, safety
and business models. Robot Comput Integr Manuf 81:102484. https://doi.org/10.1016/j.rcim.
2022.102484
28. Malik AA, Bilberg A (2019) Complexity-based task allocation in human-robot collaborative
assembly. Ind Robot 46:471–480. https://doi.org/10.1108/IR-11-2018-0231
29. Malik AA, Bilberg A (2019) Developing a reference model for human–robot interaction. Int J
Interact Des Manuf 13:1541–1547. https://doi.org/10.1007/s12008-019-00591-6
30. Rahman SMM, Sadrfaridpour B, Wang Y, Walker ID, Mears L, Pak R, Remy S (2020) Trust-
triggered robot-human handovers using flexible manufacturing, 1–8
31. Edward J, Wannasuphoprasit CW, Peshkin MA (1996) Cobots: robots for collaboration with
human operators. In: International mechanical engineering congress and exposition, Atlanta
32. Krüger J, Lien TK, Verl A (2009) Cooperation of human and machines in assembly lines. CIRP
Ann Manuf Technol 58:628–646
33. Malik AA, Bilberg A (2019) Human centered lean automation in assembly. Procedia CIRP
81:659–664. https://doi.org/10.1016/j.procir.2019.03.172
34. Ajoudani A, Zanchettin AM, Ivaldi S, Albu-Schäffer A, Kosuge K, Khatib O (2018) Progress
and prospects of the human–robot collaboration. Auton Robots 42:957–975. https://doi.org/
10.1007/s10514-017-9677-2
35. Povlsen K (2023) A new generation of robots can help small manufacturers. https://hbr.org/
2023/11/a-new-generation-of-robots-can-help-small-manufacturers
332 A. A. Malik et al.

36. BMW (2017) HRC system in production at BMW Dingolfing

37. Bänziger T, Kunz A, Wegener K (2017) A library of skills and behaviors for smart mobile
assistant robots in automotive assembly lines. In: Proceedings of the companion of the 2017
ACM/IEEE international conference on human-robot interaction, pp 77–78
38. Kirova I (2020) Give me a hand—The potential of mobile assistive robots in automotive logistics
and assembly applications Stefanie. Educ Technol J 11:329–331. https://doi.org/10.26883/
2010.202.2349
39. Melissa R (2023) Articulated robot market size. https://statzon.com/insights/global-articulated-
robot-market
40. Malik AA, Masood T (2022) Mobile robot assistants in wind turbines manufacturing. SSRN
Electron J. https://doi.org/10.2139/ssrn.4241697
41. ISO: ISO 19649_2017(en) Mobile robots—Vocabulary
42. Grieves M (2002) Completing the cycle: using PLM information in the sales and service
functions, Troy
43. De Oliveira Hansen JP, Da Silva ER, Bilberg A, Bro C (2021) Design and development of
automation equipment based on digital twins and virtual commissioning. Procedia CIRP.
104:1167–1172. https://doi.org/10.1016/j.procir.2021.11.196
44. Grieves M, Vickers J (2017) Digital twin: Mitigating unpredictable, undesirable emergent
behavior in complex systems. In: Transdisciplinary perspectives on complex systems. Springer,
pp 85–113
45. ISO: ISO 10303-242_2022—Industrial automation systems and integration—Product data
representation and exchange—Part 242_Application protocol_Managed model-based 3D
engineering
46. International Electrotechnical Commissioning (2003) Programmable controllers-Part 1:
General information
47. Malik AA, Brem A (2021) Digital twins for collaborative robots: a case study in human-robot
interaction. Robot Comput Integr Manuf. 68. https://doi.org/10.1016/j.rcim.2020.102092
48. Qi Q, Tao F, Hu T, Anwer N, Liu A, Wei Y, Wang L, Nee AYC (2021) Enabling technologies
and tools for digital twin. J Manuf Syst 58:3–21. https://doi.org/10.1016/j.jmsy.2019.10.001
49. Collaborative Robots (2016) Eliminate the mystery around safety robot safety standards for
collaborative operation ISO 10218 2016. Curr Saf Stan
50. ISO: ISO 13849-1_(2023)(en) Safety of machinery—Safety related parts of control systems,
Part 1: General principles for design
51. Robots ISO (2016) Robotic devices--collaborative robots. ISO 15066: 2016
52. Bauer W, Bender M, Braun M, Rally P, Scholtz O. Lightweight robots in manual assembly—
best to start simply! In: Examining companies’ initial experiences with lightweight robots,.
Frauenhofer-Institut für Arbeitswirtschaft und Organisation IAO, Stuttgart
53. Tan JTC, Duan F, Zhang Y, Watanabe K, Kato R, Arai T (2009) Human-robot collaboration
in cellular manufacturing: design and development. In: intelligent robots and systems, 2009.
IROS 2009. IEEE/RSJ International Conference on. pp. 29–34 (2009)
54. Francalanza E, Borg J, Constantinescu Deriving a systematic approach to changeable
manufacturing system design. In: Procedia CIRP. pp 166–171. Elsevier B.V.
55. Malik AA, Bilberg A (2017) Framework to implement collaborative robots in manual
assembly: a lean automation approach. In: Katalinic B. (Ed.) Proceedings of the 28th DAAAM
international symposium, Published by DAAAM International, ISSN. pp 1726–9679
56. ElMaraghy H (2014) Managing variety in manufacturing. Procedia CIRP 1–2 (2014)
57. ROS: Robot Operating System, https://www.ros.org/
58. GitHub: OpenPLC_ an open source industrial controller. https://github.com/thiagoralves/Ope
nPLC
59. EPIC Games: Unreal Engine. https://www.unrealengine.com/en-US
60. IEC: IEC 63278-1 (2023) ED1 Asset Administration shell for industrial applications—Part 1:
Asset administration shell structure. (2024)
Mobiliti: A Digital Twin for Regional
Transportation Network Design
and Evaluation

Jane F. Macfarlane and Ismaeel Babur

Abstract Mobiliti is introduced as a foundation for a Digital Twin for managing

transportation systems across metropolitan regions and for planning and designing
urban networks. Regional transportation systems consist of interconnected subnet-
works, each governed by different municipalities. Although localization simplifies
analysis, transportation projects must be evaluated within the context of the larger
regional network due to their potential impacts on overall network performance. The
computational challenges of simulating metropolitan networks are addressed through
parallel discrete event simulation, enhancing the predictive capabilities of the Digital
Twin framework. This technology demonstrates significant potential to enhance
urban quality of life, encourage shifts to multi-person vehicle use, and provide action-
able feedback on infrastructure investments to ensure that urban environments are
economically viable, safe, and healthy.

Keywords Regional-scale transportation dynamics · Parallel discrete event

simulation, · Traffic assignment · Equity measures · Road networks · Real-time
mobile device data · Predicitive traffic · Generative network design · Greenfield
city design

1 Introduction

The US transportation sector is in the midst of a significant digital transformation

[12]. An increasing number of connected vehicles are appearing on our roads as orig-
inal equipment manufacturers (OEMs) integrate advanced sensors into cars, trucks,
and public transportation. These sensors, which include cameras, lidar, GPS, and

J. F. Macfarlane (B) · I. Babur

Smart Cities Research Center, Institute of Transportation Studies, University of California
Berkeley, 115 McLaughlin Hall Berkeley, Berkeley, CA 94720, USA
e-mail: [email protected]
I. Babur
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 333
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_14
334 J. F. Macfarlane and I. Babur

radio communication devices, are collecting vast amounts of data on vehicle loca-
tion, speed, surrounding environments, and the health of components. This data is
then analyzed to derive insights into vehicle dynamics, such as driving patterns,
congestion levels, and critical braking events.
At the same time, the infrastructure itself is becoming smarter. Sensors and
cameras installed across various points enable remote monitoring of traffic condi-
tions, asset status, accidents, and weather changes. Traffic management centers are
using this wealth of information to make operational decisions that enhance response
times and overall efficiency. This shift is replacing older hardware systems with intel-
ligent transportation systems (ITS), which include automated features like adaptive
traffic signals, dynamic tolling, and smart parking meters, all designed to streamline
transportation management.
Additionally, the surge in data is being captured and analyzed through big data
platforms, hosted on the cloud. These platforms aim to provide a holistic view of
transportation networks by integrating data from both vehicles and infrastructure
sensors.
A transformation toward Mobility-as-a-Service (MaaS) is taking place. This
model leverages the new digital and communication infrastructure to offer on-
demand services such as ride-sharing, bike-sharing, and microtransit through
consumer-friendly mobile apps. These services are becoming viable alternatives
to private car ownership, particularly in urban areas. Alongside this shift, there is
a growing governmental push toward electrification, including the development of
charging networks and electric fleets, aimed at mitigating the environmental impacts
of internal combustion engines. Digital systems are also playing a crucial role in
facilitating the integration and monitoring of these new electric vehicle demands.
As a whole, this evolving ecosystem of digital technologies is not only enabling
new transportation services and optimizing infrastructure usage, it is also enhancing
the capacity for comprehensive monitoring and data-driven planning and operations.
This transformation is set to accelerate and will reshape the future of transportation
in the years to come.

2 Digital Twins for Regional Transportation Networks

Urban transportation networks are inherently complex, consisting of numerous inter-

connected systems and a variety of stakeholders, both public and private. To navigate
and influence this intricate landscape effectively, all involved entities must deeply
understand the dynamic interactions associated with the movement of people and
goods. Traditional static models and simulations have often fallen short, limited by
their inability to fully capture the intricacies of regional transportation dynamics.
Emerging and evolving sensors that monitor transportation systems are paving
the way for innovative solutions like Digital Twins. These tools provide a virtual
representation of transportation systems, aligning real-time data with system models
to mirror the physical network’s key components. This allows planners to simulate
Mobiliti: A Digital Twin for Regional Transportation Network Design … 335

various scenarios, understand the cascading effects of network changes, and impor-
tantly, identify unintended consequences of these modifications. Enhanced compu-
tational power and advanced simulation technologies are amplifying the capabilities
of Digital Twins, offering unprecedented insight into the dynamics of transportation.
The development of effective Digital Twins for regional planning hinges on several
key research areas: defining clear fundamental domain taxonomies, advancing simu-
lation methodologies, integrating data from multimodal transportation sources, and
using artificial intelligence to codify urban complexities. Thoughtfully designed,
Digital Twins have the potential to revolutionize the transportation planning process.
They enable more precise analyses of investment options and policy decisions, offer
the capability to evaluate alternatives swiftly, and provide rapid feedback, all of
which contribute to more agile and effective decision-making processes that enhance
mobility outcomes. Realizing the full potential of Digital Twins will necessitate
strong partnerships among researchers, technology providers, planning agencies,
and private industry.

2.1 Modeling Regional Transportation Systems

Regional transportation systems consist of complex networks made up of intercon-

nected subnetworks, each often governed independently by different municipalities.
Although local governance can simplify some aspects of analysis, it is crucial for
transportation projects to consider their role within the broader regional context due to
significant impacts on network performance. This is particularly important as urban-
ization continues to increase vehicle densities in metropolitan areas. Historically,
the capability to analyze extensive regional networks was hampered by computa-
tional limitations, necessitating substantial simplifications in network and demand
modeling.
Metropolitan regions comprise multiple cities, each with its unique dynamics
influenced by its geographical setting. Cities typically operate autonomously, and
smaller cities may be particularly constrained by limited financial resources. This
autonomy often results in a patchwork of traffic management strategies, linked
together only by state Department of Transportation (DOT) managed interstates and
highways.
Within cities, traffic management is handled through various mechanisms
including traffic regulation tools like signals and stop signs, road network controls
such as speed limits and turn restrictions, physical traffic calming measures like speed
bumps and lane narrowing, and infrastructure investments focused on non-vehicular
traffic, including protected bike lanes and intersection curb extensions. Meanwhile,
state DOTs are responsible for managing broader infrastructure elements like inter-
states, state highways, and bridges, overseeing broader traffic sensing, incident
management, and ramp metering.
Adding another layer to this complex system, navigation applications in vehicle
dashboards and smartphones act as a third party influencing metropolitan traffic
336 J. F. Macfarlane and I. Babur

management. These apps can dynamically reroute travelers to faster routes as they
become available, which, while while potentially helpful in reducing congestion, can
also lead to unintended traffic disruptions in residential areas or lead to ill-informed
behaviors in unusual circumstances, such as during wildfire evacuations.
The primary objective of regional transportation modeling is to codify these
complexities of disjointed traffic management systems. The aim is to develop
analytics that assist city planners in enhancing transportation options and dynamics
across cities and to design strategies that encourage mode shifts which will be crucial
for the future of traffic management in metropolitan areas.

2.2 Mobiliti: Transformative for Transportation Planning

Simulation tools for transportation systems come in various forms, typically catego-
rized into microsimulation, mesosimulation, and macrosimulation. Our novel plat-
form, Mobiliti, stands out by overcoming traditional computational limitations asso-
ciated with the simulation of large regional road networks. It uses an agent-based,
parallel discrete event simulation framework where each road link acts as an agent.
These agents control the movement of vehicles across their geospatial locations,
managing traffic flow according to constraints such as vehicle capacity, speed limits,
and queue lengths. For instance, in the San Francisco Bay Area, Mobiliti can model
24 h of activity for approximately 7 million personal vehicles and 4 million trucks,
translating to approximately 19 million trips across a network of 1 million road links,
in less than eight minutes—a scale of simulation and computational speed unmatched
by any current system.
Similar simulation tools limited by computational constraints are forced to reduce
the scale of the simulation. This typically involves decreasing the number of trips
or population size, reducing the size of the road network or simplifying its geom-
etry, or any combination of these approaches. These reductions and simplifications
fail to accurately represent the true dynamics of the metropolitan regions being
studied. In practice, this limitation is particularly critical when considering dynamic
rerouting, which often involves lower-class neighborhood roads that are often omitted
in reduced network models.
In contrast, planning agencies sometimes use optimization algorithms to under-
stand road network loading. These tools analyze trip demand profiles and assign
trips to the network, aiming to optimize travel times. However, these tools also face
computational challenges, and the traffic assignment models often operate with very
low spatial and temporal fidelity—for example, running optimizations over a static
four-hour period. Despite this, these tools can still require up to 24 h to process and
frequently yield results that significantly underestimate congestion patterns.
Mobiliti: A Digital Twin for Regional Transportation Network Design … 337

2.3 The Need for Regional Models

The importance of regional transportation models stems from their ability to accu-
rately represent the travel demand profiles of individual cities, a necessity for several
reasons. First, economic and land use impacts are inherently regional. Activities such
as employment, shopping, and recreation generate travel patterns that often begin in
one city and conclude in another within the same region. Consequently, examining a
single city in isolation does not capture critical origin–destination flows and requires
assumptions about boundary conditions for trips that cross these boundaries.
Moreover, the transportation needs of a city are deeply influenced by regional
developments. Public transportation options and road networks extend across city
limits, creating an interconnected system where traffic conditions in one area
can affect another downstream. Additionally, cities often share significant infras-
tructure such as ports, airports, and major state-run corridors, emphasizing the
interdependence of urban areas.
Transportation policy and investment decisions, typically made at a regional level,
require insights into travel behaviors across multiple cities. By focusing solely on
individual cities, significant interactions and the use of shared assets may be over-
looked. Regional models address these challenges by providing a comprehensive
view of mobility patterns, capturing the intricacies of how people move within and
between urban areas.
With projections indicating that 68% of the global population will reside in major
cities by 2050, the relevance of regional models becomes even more pronounced. As
urbanization continues to drive growth in major cities and expands them into future
mega-regions, understanding the complex dynamics of regional transportation will
be crucial for managing congestion and planning sustainable urban environments.

3 Digital Twin Versus Model Based Controller

When discussing Digital Twins, it’s important to recognize that they are composed
of three distinct components: a physical system operating in real-time, dynamic data
about this system streamed at various frequencies, and a computationally manage-
able simulation of the system. This chapter explores a Digital Twin specifically
designed to address the complex challenges of managing transportation systems
across metropolitan regions.
The Digital Twin Consortium defines a Digital Twin as “a virtual representa-
tion of real-world entities and processes, synchronized at a specified frequency and
fidelity”. This emphasizes the Digital Twin’s role in mirroring real-world systems
and processes accurately and in real-time.
Contrasting with this is the model-based controller, a concept well-established
in traditional control theory. Model-based controllers use knowledge of physical
laws governing systems to estimate responses to external inputs. This knowledge is
338 J. F. Macfarlane and I. Babur

embedded within a feedback control system, allowing the controller to predict and
optimize system behavior under various conditions. The challenge lies in developing
accurate models that can sufficiently account for all relevant system behaviors.
In comparison, a Digital Twin goes beyond control. It serves as a comprehensive
virtual counterpart of the system, not only monitoring and optimizing performance
but also enabling scenario testing without impacting the physical system. Digital
Twins thus offer substantial benefits, such as providing valuable insights, facilitating
remote monitoring and control, and enhancing decision-making processes.
Key differences between a model-based controller and a Digital Twin include:
. A model-based controller uses a simplified model primarily for control, aiming to
meet specific control requirements. It does not typically mirror the full complexity
of the physical system. Conversely, a Digital Twin seeks to emulate the physical
system as closely as possible, incorporating much greater detail and complexity.
. The model in a model-based controller is generally static, with updates based on
new data being slow or non-existent. On the other hand, a Digital Twin contin-
uously updates its virtual representation with real-time data from the physical
system.
. The primary objective of a model-based controller is to calculate optimal control
inputs to achieve desired performance. In contrast, the Digital Twin aims to
understand, analyze, and potentially optimize the physical system, providing a
predictive analytics capability for exploring future states.
While model-based controllers use simplified models designed specifically for
control algorithms, Digital Twins aim to fully virtualize the physical system and
maintain synchronization with real-time data. This allows Digital Twins to facilitate
a broader exploration of scenarios in the context of the system’s current state. It is
likely that as Digital Twins evolve, model-based control techniques will be integrated
within them, validated against these higher-fidelity representations to further enhance
their utility and effectiveness.

4 Managing Computational Complexity

Transportation modeling can become exceedingly complex at larger scales. To

manage this complexity, a key feature of any Digital Twin for transportation
networks is its computationally manageable simulation capability. As the advance-
ments predicted by Moore’s law began to plateau, there was a significant shift in
focus toward developing new approaches for traditional algorithms, particularly in
the realm of parallel computation.
Parallel discrete event simulation [10] has emerged as a pivotal technology in this
context. It offers a new level of scalability for modeling extensive network systems.
Mobiliti uses this simulation approach to tackle the challenges posed by regional-
scale transportation networks, which are increasingly common in urban metropolitan
areas. This platform leverages parallel discrete event simulation to effectively manage
Mobiliti: A Digital Twin for Regional Transportation Network Design … 339

and simulate the complex dynamics of these extensive networks, providing a valuable
tool for understanding and optimizing transportation infrastructure.

4.1 Parallel Discrete Event Simulation

Parallel discrete event simulation (PDES) is a specialized simulation technique that

represents processes as a sequence of time-stamped events occurring on predefined
actors. These actors use local state information to schedule subsequent events. A
prime example of PDES application is in vehicle transportation simulation, the focus
of the Mobiliti development effort. In this scenario, the road network is divided into
segments or links, which serve as the actors, while vehicles are modeled as the events
that move between these links. The vehicle travel times are predicted based on delays
caused by the presence of other vehicles on the same link.
One of the significant challenges in discrete event simulations is the effective
parallelization of these processes. The core promise of a PDES engine is that by the
simulation’s conclusion, all events will have been processed in an order that respects
the chronology of their timestamps. In context of the complexity of transportation
modeling, where predicting event arrival times on any given actor is inherently
challenging, the PDES community developed two main strategies for parallelization:
. Conservative Parallelization: This method uses model-specific information to
ensure that no events will unexpectedly arrive within a certain time frame, thus
guaranteeing that the simulation within that time interval is safe to execute.
Although this approach often results in reduced parallelism, it can achieve good
performance if sufficient detail about the problem is known to expose enough
parallel opportunities.
. Optimistic Synchronization: Unlike conservative methods, optimistic synchro-
nization does not require that events are initially executed in the correct order.
Instead, it allows events that are processed out of order to be replayed to ensure
that the final state of each actor aligns with what would have occurred had all events
been executed correctly. However, this approach can face challenges if too many
events are executed incorrectly, leading to significant runtime overheads. Addi-
tionally, making an event “reversible” often involves considerable coding effort
and can present substantial engineering challenges, especially when third-party
libraries are used.
Deciding between these two parallelization techniques involves understanding
their performance trade-offs, which can vary widely depending on the specific
characteristics of the problem at hand. Conservative methods generally offer more
predictable performance but at the cost of limited scalability, while optimistic
approaches provide greater flexibility but can lead to increased complexity and
potential performance penalties.
340 J. F. Macfarlane and I. Babur

4.2 Discrete Event Simulation

Discrete event simulation (DES) is a computational approach used to model systems

that evolve over time through instantaneous state changes, known as events, which
occur at specific timestamps. Each event not only modifies the state of the system
but also has the potential to trigger additional future events. A DES application is
characterized by its state space, an initial state, and an initial set of events. The
framework responsible for DES ensures that these events are executed in the order
of their timestamps, continuing until no further events remain, at which point the
simulation concludes.
To manage the sequence of events effectively, DES typically employs a time-sorted
queue. This queue continuously checks for the event with the earliest timestamp and
processes it until the queue is empty. Building upon this, parallel discrete event simu-
lation (PDES) extends DES by introducing mechanisms that allow for the parallel
processing of events. This is achieved by dividing the application state into logical
processes (LPs), with each LP handling events that only affect its specific portion
of the state. This segmentation enables safe concurrent execution of events across
different LPs.
However, parallelizing events introduces significant complexity, particularly in
maintaining the correct order of event execution within each LP. A typical challenge
arises when events from different LPs are executed concurrently, and an event from
one LP triggers a new event in another LP with a timestamp that predates the currently
processed event. This scenario leads to a violation of the time-ordering constraint
and creates a conflict in the simulation’s execution process.

4.3 Conservative Versus Optimistic Execution

To manage these complexities, two general strategies are employed: conservative

and optimistic execution.
In the conservative approach to execution, events within a runtime are only
processed when it is certain that no preceding events will be generated. Typically, this
means that the only event deemed safe for execution is the one with the earliest global
timestamp. This method, however, imposes a strict global linear order of event times,
effectively limiting opportunities for parallel computation and thus constraining the
computational scalability.
To optimize this approach, events can be distributed across multiple CPUs by
dividing the master event queue into several per-LP (logical process) queues. The
application must provide the minimum timestamp difference for any pair of events
running in respective LPs, where the first event triggers the second. Events are consid-
ered safe for concurrent execution if their timestamps are within the bounds set
by their respective LPs in this matrix. During execution, CPUs must frequently
exchange their earliest timestamp values to calculate the most lenient execution
Mobiliti: A Digital Twin for Regional Transportation Network Design … 341

bounds. However, due to the necessity of adhering to the most conservative case,
this method often results in missed opportunities for parallel execution and requires
aggressive synchronization.

4.3.1 Optimistic Execution

Optimistic execution, pioneered by Jefferson in his description of the Time Warp

mechanism [7], adopts a more localized approach. It operates under the assumption
that the earliest event in an LP’s queue is likely safe to execute, despite the risk that
an earlier event might arrive from another LP. To manage this risk, the runtime must
be capable of ‘rolling back’—reverting the state of the LP to the point before the
out-of-order event was executed. This process requires that the application provides
mechanisms to reverse one or more events, a complex challenge in itself. Addition-
ally, the system must handle potentially invalid event orderings, which will need
to be corrected through rollbacks. The key advantage of optimistic execution over
conservative methods is that it allows for the execution of events without immediate
synchronization, taking advantage of the frequent safety of event execution despite
the lack of concrete proof of valid event ordering.

4.3.2 Asynchronous Global Virtual Time Computation

A crucial aspect of any Time Warp-based parallel discrete event simulation (PDES)
is managing the computation of Global Virtual Time (GVT). GVT is the consensus
timestamp up to which all events are confirmed as correctly processed, serving as a
baseline for system integrity. Maintaining a queue that can revert the system to any
prior state is essential for enabling rollbacks when events are processed out of order,
though this can significantly increase memory requirements.
To enhance scalability and manage this complexity, a highly asynchronous version
of Time Warp is implemented that eschews blocking collective operations. A primary
challenge in optimizing Time Warp performance is to curb excessive optimism, which
can lead to frequent and resource-intensive rollbacks. Rollbacks are necessitated
when an event, once processed, must be reversed due to subsequent receipt of an
earlier time-stamped event. This also involves rolling back any events that the initial
event might have triggered on neighboring processors.
The system leveraged by Mobiliti, called Devastator, was developed to run on a
supercomputer at Lawrence Berkeley National Laboratory by Cy Chan, Max Bremer,
et al. It employs a strategy known as the moving time window. This method limits
the execution interval to GVT plus a tunable parameter (w), thus containing the
temporal range within which events can be executed. As GVT progresses, this
window advances, progressively allowing more pending events to be processed.
Devastator automatically adjusts the width of this window based on operational
efficiency metrics.
342 J. F. Macfarlane and I. Babur

In summary, while the conservative synchronization approach grapples with the

scalability challenges posed by frequent collective synchronization, the optimistic
synchronization method offers a solution. By reducing the latency associated with
smaller, point-to-point messages, it decreases the likelihood of delays caused by
‘straggler’ events that necessitate rollbacks, thus enhancing the system’s scalability
and suitability for application needs.

5 Mobiliti—A Platform for Regional Traffic Modeling

Agent-based traffic simulation serves as a robust mechanism for understanding real-

world transportation systems and facilitating the development of Digital Twins [16,
17]. By leveraging Devastor, Mobiliti is designed to effectively model extensive
regional transportation networks. This is achieved by treating road links as parallel
actors that manage the flow of vehicle events. Mobiliti is structured upon three
primary software layers: GASNet-Ex, Devastator, and Mobiliti itself. GASNet-Ex [5]
is an open-source software that facilitates high-performance inter-process communi-
cations across distributed memory systems, ideal for handling small active messages.
Devastator operates above GASNet-Ex, implementing the Time Warp optimistic
parallel discrete event protocol which manages the scheduling, execution, rollback,
commit, and Global Virtual Time of events. At the top layer, Mobiliti introduces
domain-specific logic that defines both the actors and events within the road network,
and it orchestrates the parallel domain decomposition, mapping actors to processors.
Within Mobiliti, road links act as computational actors, calculating vehicle
traversal times and enforcing timing and capacity constraints, which naturally lead to
queue formation as congestion builds. The platform tracks congestion by monitoring
vehicle flow rates on these links across specified time windows. Designed to run
on the supercomputer at Lawrence Berkeley National Laboratory, Mobiliti demon-
strates strong scalability, effectively running on up to 1024 cores. Performance tests
reveal that the baseline simulation can process over 2.4 billion discrete events in less
than 30 s using 1024 cores, handling massive networks efficiently [2].
Beyond simulating basic vehicle movement across a network, Mobiliti also
supports dynamic rerouting, allowing a predetermined number of vehicles to request
new routes based on real-time congestion data. This feature reflects the pervasive
use of navigation systems in modern urban settings and offers adjustable controls for
rerouting parameters and the distribution of rerouting permissions among vehicles.
Figure 1 outlines the components of Mobiliti illustrating how different data inputs
interact and contribute to the overall modeling process. The model is broken down
into three primary components, each colored differently to represent distinct aspects
of the simulation:

1. Road Network (Green): Represents the physical layout of the transportation

infrastructure, including all roads and paths vehicles can traverse. It provides the
foundation for the simulation where all other data will interact.
Mobiliti: A Digital Twin for Regional Transportation Network Design … 343

Fig. 1 Overview of Mobiliti highlighting the interconnected components used to analyze and
optimize urban mobility

2. Trip Demand (Yellow):

. ODT with Geodemographics: Origin–destination time data enhanced with
geodemographic information from a travel demand model helps in under-
standing the demographic characteristics of travelers and their travel patterns.
. Truck Model: Specific modeling considerations for truck movements, which
might differ in terms of routes, regulations, or times of operation compared
to other vehicles.
3. Fleet Composition (Gray): Details the types of vehicles involved in the
transportation network and is used for determining energy use.
The outputs of the model, based on the inputs from the three areas above, are
divided into three types of metrics:
. Link Metrics: Measures related to individual road segments such as flow, speed,
delay, and energy consumption.
. Trip Metrics: Focuses on the efficiency and impact of individual trips within the
network, including travel time, delays, and energy usage.
. Routes: Involves both static and dynamic routing information. Static routes are
predetermined, while dynamic routes can change in response to traffic conditions,
blockages, or other real-time factors.
The model uses these inputs and outputs to simulate and analyze transportation
scenarios, optimizing for various factors such as reduced travel times, decreased
energy consumption, or improved safety.
Mobiliti uses road network graphs derived from HERE Technologies maps [6] or
may use any detailed road network as long as essential attributes like free-flow speed,
lane count, and road geometry are specified. An example network built using open-
source data for Houston is detailed in [4]. The vehicle demand profile is typically
generated from regional travel demand models and specifies the origin, destination,
and timing of trips.
344 J. F. Macfarlane and I. Babur

Mobiliti employs an arc-based routing algorithm that modifies Dijkstra’s classic

shortest path algorithm to operate directly on road segments rather than intersec-
tions. This approach simplifies the graph representation and improves cost modeling
accuracy by associating costs like travel time or fuel consumption directly with road
segments. It avoids the need for node preprocessing and sidesteps the topological
constraints required by traditional Dijkstra’s algorithm, allowing for more flexible
and efficient routing. Connected car data was used to determine cost modeling and
is an example of how the Digital Twin is enhanced by integrating real-world data.
A simulation of the Los Angeles area, visualized using kepler.gl in Fig. 2, demon-
strates the platform’s capabilities on a network of approximately 2 million links
with around 40 million trips in a single day, completing the simulation in about
15 min. This simulation was based on data from the 2020 Southern California Area
Government Group Travel Demand Model. Validation of simulation results incor-
porates comparisons against multiple real-world data sources such as PeMS, Uber
Movement speeds, on-road tube counts, regional surveys, and products from large
real-time data collection vendors. Mobiliti has recently surpassed the regional level
by demonstrating its ability to handle simulations at an unprecedented scale. The
California Statewide Travel Demand Model was used to simulate vehicle and truck
movements across the entire state, accounting for over 102 million trips. This repre-
sents one of the largest transportation simulations ever created, offering new insights
into statewide traffic dynamics and congestion patterns. These statewide insights are
crucial, as regional and city-level simulations often struggle to adequately represent
‘out-of-region’ or ‘out-of-city’ trips, frequently resorting to assumptions or simpli-
fications that fail to capture the full scope of traffic flows beyond their planning
jurisdictions. Figure 3 visualizes the daily vehicle count over the entire California
road network, based on results from the statewide simulation.

6 Enriching Digital Twins Through Advanced Data

Integration

Developing a precise and operationally effective Digital Twin requires a detailed,

high-quality depiction of the dynamics and behaviors of real-world transportation
systems. This involves modeling and quantifying the complex interactions among
various components and features throughout the regional network. The model’s
complexity is further enriched by integrating dedicated sensor data and ancillary
datasets, enhancing its ability to account for unpredictable events, such as traffic
incidents or unplanned construction on routing behaviors.
Mobiliti: A Digital Twin for Regional Transportation Network Design … 345

Fig. 2 Mobiliti link speeds for the Los Angeles basin using the 2020 SCAG travel demand model,
visualization by kepler.gl

6.1 Overcoming Data Fragmentation

Traditionally, government entities have developed or acquired software solutions

tailored to specific operational needs, resulting in a fragmented data landscape
where information is optimized for narrow applications within each department. For
instance, individual departments may often collect data primarily to analyze conges-
tion patterns to support transit planning or infrastructure investments. However, this
domain-specific data is not easily usable or accessible to other city stakeholders who
might leverage this traffic analysis to identify logistics inefficiencies or plan economic
development initiatives. This underscores the importance of an interoperable Digital
Twin framework that dismantles data silos and fosters a shared understanding across
various urban applications.
While some government agencies have begun to use emerging data sources and
analytical methods, these ad hoc approaches frequently lead to gaps and inefficien-
cies. Specific engagements with outside services often generate costly static reports
with limited analytical capabilities, and the data remains siloed. Despite substantial
investments in data collection, the lack of interoperability restricts the broader appli-
cational use of these datasets, leaving their full potential value untapped. By orga-
nizing heterogeneous data into a unified, interoperable knowledge graph, datasets
346 J. F. Macfarlane and I. Babur

Fig. 3 CA statewide daily vehicle count

originally intended for single purposes can be transformed into a comprehensive

digital representation, supporting a broad spectrum of smart city functions.
A robust Digital Twin framework must include a common knowledge representa-
tion and adapters to align diverse data models into an extensible ontology. This facili-
tates the portability of analytics, allowing data science workflows and AI/ML models
to be applied broadly within the Digital Twin, extracting reliable and reproducible
insights from the integrated data set.
Larger municipal agencies might develop proprietary in-house solutions or engage
technology firms to create custom systems. However, this approach often forces
government bodies to assume the roles of software developers, taking on significant
risks and costs without necessarily having the needed expertise. Conversely, the
Mobiliti: A Digital Twin for Regional Transportation Network Design … 347

growing trend of using singular focused applications developed by tech companies

offers proven solutions but still leads to siloed applications within municipalities.
To move beyond these challenges, a holistic Digital Twin framework is essen-
tial. It should provide a unified data infrastructure and an integrated environment for
data and model integration, allowing seamless AI/ML workflows to operate across a
consolidated, multi-source knowledge graph. This framework enables the develop-
ment of reusable domain-adapted models that can be transferred across the Digital
Twin’s applications through shared service abstractions.
In contrast to custom, siloed applications, a model-driven Digital Twin archi-
tecture separates domain logic from data structures, enhancing portability and mini-
mizing the need for rework. Versioned analytics and model management are essential
for supporting robust tracking of assets and data lineage throughout the Digital Twin
ecosystem, ensuring consistent and reliable decision support.

6.2 Digital Twin Data Infrastructure

The ultimate aim is to establish an explorative platform built upon unified data that
represents a single source of truth while enabling Digital Twin scenario development
capabilities. APIs that facilitate seamless access to this curated data and secure inte-
gration of third-party datasets will encourage users to contribute back to the collective
pool. This fosters a beneficial cycle of data sharing and enhancement across domains,
agencies, and jurisdictions.
Integrated data science tools embedded within the platform can extract reliable,
reproducible insights from the combined data, positioning the Digital Twin as the
authoritative source for interdisciplinary analysis and preventing the loss of insti-
tutional knowledge. By maximizing the value of multi-source data fusion, a shared
data infrastructure enhances both data quality and reliability through scalable data
science techniques.
For example, predictive models trained on extensive historical data can detect
anomalies in incoming data streams that deviate from expected patterns. These
conflicts among data from different sensor types can be automatically identified and
flagged for review. This multi-source data fusion can potentially mitigate inherent
biases in individual datasets, revealing biases related to underrepresented populations
and usage patterns, thus supporting more equitable decision-making.
These capabilities, integrated into an automated data observatory that contin-
uously monitors the Digital Twin’s data ecosystem and identifies accuracy issues
through AI/ML validation pipelines, are crucial to the Digital Twin’s performance.
Issues are prioritized based on their impact, with notifications sent to relevant experts
for necessary adjustments. Engaging the relevant experts in the organization and
involving them in model adjustments based on trending data will be essential to the
long-term success of the Digital Twin.
Given baseline data quality is verified, completeness can be addressed. For the
transportation domain, data coverage varies widely across different communities and
348 J. F. Macfarlane and I. Babur

may be sparse in certain areas or during specific time periods. Figure 4, for instance,
illustrates traffic count data collected by contractors over several years in San Jose.
In situations where data gaps exist due to limitations in spatial or temporal
coverage, or sensor malfunctions, the comprehensive data repository of the Digital
Twin can be employed to create synthetic data to fill these gaps. Deep learning
techniques like graph neural networks can assist in this case [14, 15].
However, it is critical to quantify the uncertainty associated with these synthesized
data products. Real-time data can instill confidence by appearing to accurately reflect
reality, yet it is very often subject to anomalies due to hardware issues, collection
methods, and varying penetration rates. By establishing confidence intervals and
error propagation models, downstream users can assess the reliability of the data
and make informed decisions regarding its suitability for their workflows. This is
illustrated in Figs. 5 and 6, which showcases the data validation processes using
Mobiliti results.
These advanced algorithms provide a detailed and comprehensive view of regional
behaviors, integrating various influencing factors such as time of day, seasonal varia-
tions, nearby events, weather conditions, population densities, the built environment,

Fig. 4 Sparsity of static sensor data

Mobiliti: A Digital Twin for Regional Transportation Network Design … 349

Fig. 5 Validation of Mobiliti results with static sensor data

Fig. 6 Validation with dynamic sensor data with uncertainty

and the economic landscape. These elements are combined within the Digital Twin
to predict traffic flows, speeds, and congestion patterns at a granular level across the
transportation network.
The simulation model allows for adjustments in individual parameters to gauge
the sensitivity of output metrics to specific changes in the input conditions. The
accuracy of the simulation engine is validated by comparing its predicted outputs
with real-world ground truth data within the Digital Twin demonstrating a favorable
350 J. F. Macfarlane and I. Babur

Fig. 7 Vehicle count validation allows the Digital Twin to estimate vehicle counts (left) and speeds
(right) across the full network—colored from red to yellow

alignment between simulated and observed data. Once the simulation fidelity is
established through this rigorous benchmarking, the calibrated engine can be used
to fill missing data gaps by generating synthetically simulated values, as shown in
Fig. 7. This approach, corroborated against available estimates, ensures a continuous,
holistic view of the transportation system’s state.
For short-term predictions, the simulation engine can model the impacts of vari-
ables such as event attendance, weather forecasts, or disruptions in transit services
on expected traffic conditions. For longer-term planning and policy analysis, it eval-
uates scenarios related to demographic shifts, changes in land use, or housing devel-
opments to estimate their effects on key performance indicators including mobility,
accessibility, equity, safety, and environmental impacts.
The interplay between data-driven simulation and the central unified data repos-
itory enhances the Digital Twin framework’s capabilities. The simulation engine
continuously improves its accuracy by assimilating the latest real-world observa-
tions and can also augment the data repository in areas showing deficiencies. This
closed-loop data enrichment ensures that the Digital Twin maintains a comprehen-
sive, self-regulating representation of the physical system’s state , enabling informed
and reliable decision-making.
An integrated multi-source data environment facilitates the discovery of novel
relationships and insights through advanced analytics and machine learning. By
applying deep learning techniques to the consolidated data corpus, previously hidden
patterns and complex correlations are uncovered, enabling more accurate modeling
and prediction of system behaviors, as highlighted in Fig. 7, where the real-time data
and learned behaviors tune the model and overcome data sparsity and penetration
concerns.
Ultimately, by dismantling long-standing data silos and establishing an integrated
Digital Twin environment, cities can unlock the immense potential of their data
assets, deriving reliable, actionable insights to drive intelligent policies, resilient
infrastructure investments, and equitable resource allocation across all communities.
Mobiliti: A Digital Twin for Regional Transportation Network Design … 351

6.3 Improving Travel Demand Estimates and Traffic Insights

Proper alignment of real-world data with simulation models is critical for building
trust in the Digital Twin [11]. Travel demand models, which are intricate and costly
to develop, are typically updated at five year intervals. One of the core advantages
of Digital Twins is their ability to incorporate real-time data to refresh these models
more frequently, reflecting the latest trends in demographics, economic conditions,
and land use.
Moreover, as the Digital Twin generates a comprehensive virtual representation of
traffic, new insights can emerge from the observed traffic patterns. Unlike traditional
methods that may overlook fine-grained details, Mobiliti enables a detailed analysis
across the entire network. For instance, critical link density, detailed in Fig. 8, is
indicative of where vehicle accumulation peaks during the day and provides valuable
insights that might not be feasible to capture through conventional methods.
Derived traffic insights highlight links that are consistently over-capacity, shed-
ding light on potential congestion causes. Figure 9 identifies locations of saturation
(left) and temporal congestion (right). This information, coupled with expert anal-
ysis, can lead to innovative infrastructure solutions or policy adjustments that alleviate
traffic issues.
The ability to simulate and predict traffic under various conditions is another
significant benefit of the Digital Twin. By adjusting simulation parameters, the
model can assess the impact of changes in traffic flow, speed, and congestion across
different times of day, seasons, or in response to specific events. These simulations
are validated against real-world data, ensuring the model’s accuracy and reliability.
Figure 10 compares Average Daily Traffic (ADT) from Mobiliti with actual
observed data, demonstrating the model’s alignment with real-world conditions. It
also shows a comparison to a popular tool that relies solely on mobile device data.
This benchmarking is crucial for confirming the simulation’s accuracy before using
it to fill data gaps or guide decision-making processes.

Fig. 8 Road link level critical density indicates the point at which accumulation of vehicles occurs
352 J. F. Macfarlane and I. Babur

Fig. 9 Saturation index for the Bay Area model (left); temporal congestion index for the Bay Area
model (right)

Fig. 10 Real-world data validation with Mobiliti results

This dynamic interplay between data-driven simulation and a unified data reposi-
tory empowers the Digital Twin to continuously learn and adapt, enhancing its ability
to provide a detailed and accurate representation of the transportation system. This
ongoing data enrichment ensures that the Digital Twin remains a vital, self-updating
tool for decision support.
Ultimately, by integrating these capabilities into a cohesive Digital Twin frame-
work, municipalities can harness their data resources more effectively, driving
smarter policies, and ensuring equitable resource distribution across all communities.
Mobiliti: A Digital Twin for Regional Transportation Network Design … 353

7 Connecting City Dynamics to City Performance

While the Digital Twin capability provides powerful new computational assets
for practitioners, it does not reduce the complexity of the network fundamen-
tals. As a consequence, defining performance indices that relate to the core needs
of a city’s population remains an important endeavor. To this end, a new frame-
work for evaluating transportation policies and investments that accounts for social
equity impacts alongside traditional mobility performance metrics has been devel-
oped called Socially Aware Evaluation Framework for Transportation—SAEF [9].
Conventional evaluations often focus narrowly on mobility and fail to sufficiently
consider social outcomes. As such, a socially aware framework can provide a post-
simulation data analysis that incorporates quantitative measures of accessibility,
exposure, and externalities experienced by different social groups, including low-
income, minority, and disabled populations. These social impact measures are inte-
grated with mobility metrics within a multi-criteria decision analysis to enable a
more holistic assessment of transportation projects and policies.

7.1 Understanding the Impacts of Traffic Dynamics on City

Level Performance Indices

The complexity of city management can lead to various urban issues that signif-
icantly affect the quality of life for residents. The Mobiliti platform is currently
geared toward understanding vehicle traffic dynamics, as traffic congestion is a
prominent concern for future city planning, with direct impacts on public health,
safety, and regional economic health. Accordingly, the platform investigates the
city-level impacts of various vehicle routing strategies to enhance traffic manage-
ment and mitigate negative consequences. Beyond traffic management, it can be
used for a variety of analyses, including alternative mode planning.
The term ‘socially-aware’ encompasses four measurement themes: neighborhood,
safety, mobility, and environment. To address socially relevant aspects of vehicle
dynamics, themes and indicators are developed beyond simple congestion measure-
ments. These themes, supported by city performance indicators from extensive liter-
ature reviews [8], identify critical factors such as accidents for safety and emissions
for environmental quality. This approach has led to the creation of 24 indicators from
an initial list of over 100, categorized into themes that reflect neighborhood quality,
safety, mobility, and environmental quality.
The framework offers insights into the trade-offs between social and mobility
objectives, aiding in decision-making processes. This methodology is designed
to support more equitable transportation planning outcomes. Fifteen performance
indices, as depicted in Fig. 11, range from positive (green) to negative (red) outcomes
for four different traffic scenarios in the City of San Jose.
354 J. F. Macfarlane and I. Babur

Fig. 11 SAEF evaluation indices for the city of San Jose under four traffic scenarios (green indi-
cates a positive outcome while red indicates a poor outcome in comparison across the four traffic
scenarios). VHD vehicle hours delayed, VMT vehicle miles traveled

The Mobiliti platform includes traditional traffic assignment algorithms that

have been parallelized [3]. Four different scenarios associated with different traffic
outcomes were evaluated and detailed in Fig. 11. The scenarios include a Baseline 60
(B60) where routes are initially determined by the shortest path and then dynamically
adjusted for 60% of the trips based on real-time congestion [1]. Other scenarios use a
Frank Wolfe algorithm that optimizes trip assignments to achieve objectives like user
equilibrium travel time, system optimal travel time, or system optimal fuel consump-
tion—identified as UET, SOT, and SOF, respectively, in Fig. 11. Evaluation of the
impacts of these scenarios reflects the complexities of traffic dynamics, indicating
no single optimal approach but rather highlighting the importance of understanding
trade-offs and exploring multiple objectives in transportation planning.
The SAEF evaluation creates indicators that can then be used to take a deeper
look into the associated traffic dynamics. Vehicle Miles Traveled (VMT) and Vehicle
Mobiliti: A Digital Twin for Regional Transportation Network Design … 355

Hours Delayed (VHD) in Equity Priority Communities (EPC) can be extracted from
the simulation results. Designations of disadvantaged communities vary across cities
and area government groups and may consider a variety of factors such as income
level, population age, and ethnicity. For these analyses, geospatial representations
of these EPC communities were used to extract link level speed and flow data and
provide VMT and VHD metrics. As observed in Fig. 12, three of the four scenarios
have poor outcomes for VHD and VMT in EPCs. Further analytics, shown in Fig. 12,
then demonstrate that a significant number of longer trips (arbitrarily selected as trips
greater than 40 min) that start and end in the city centers of San Francisco and San
Jose are using the EPC road links during their journey. These are considered “pass
through” trips, in which the vehicle has no connection to the specific neighborhood
and are short-cutting congestion by using neighborhood streets on their way to their
destination. Further analysis indicates that these trips are mostly work-related trips
or commute trips and that 60% of these trips were rerouted due to congestion, thus
incurring undue safety and health impacts on lower income communities.
While the simulation may not exactly mirror reality, it provides critical insights
into traffic phenomena, particularly those influenced by driver behaviors, which are
often challenging for traffic planners to measure [13]. This demonstrates a key value
of using Digital Twins in transportation planning.

Fig. 12 Origins and destinations of San Jose and San Francisco commute trips that pass through
lower income communities thus impacting community health and safety
356 J. F. Macfarlane and I. Babur

8 Iterative Exploration

A high-fidelity virtual representation that mimics the real system enables the safe
and rapid evaluation of scenarios that are challenging or impossible to test in reality.
Gained insights can inform investment plans and policy decisions. The ultimate aim
is to equip practitioners with tools to configure the Digital Twin to reflect poten-
tial changes in infrastructure or technology and assess their impacts on system
performance.
Examples include:
. Reducing the number of lanes on an arterial road to create space for protected bike
lanes and observing the effects on congestion and traffic dynamics influenced by
dynamic routing.
. Adjusting the Origin–Destination Time (ODT) dataset in the Digital Twin to reflect
changing conditions such as population growth and demand levels and observing
the resulting strain on network resources. For instance, in San Francisco, a simula-
tion where current transit riders were instead assigned a personal vehicle showed
the significant congestion that would result without existing transit systems
(Fig. 13-left).
. Introducing disruptive events like natural disasters or major accidents into the
Digital Twin to assess resilience and response capabilities. For example, a simu-
lated earthquake in San Jose directed drivers within a certain radius to return

Fig. 13 Example scenarios for iterative exploration: include transit riders as vehicle owners/drivers
(left) and a simple earthquake model in which vehicle drivers return home (vehicle density is
indicated from red to yellow)
Mobiliti: A Digital Twin for Regional Transportation Network Design … 357

home, dramatically affecting local traffic and also impacting main arterials far
beyond neighboring cities (Fig. 12-right).
. Using machine learning within the Digital Twin to develop predictive models.
. Engaging stakeholders to define scenarios of interest and exploring consequences
to foster an intuitive understanding of existing conditions.
While simulations within the Digital Twin may not always perfectly align with
reality, the insights they provide are invaluable. They offer a safe and efficient means
for traffic planners and other practitioners to evaluate scenarios that might be chal-
lenging or impossible to test in real life. This high-fidelity virtual representation helps
inform investment plans and policy decisions, equipping practitioners with robust
tools to adjust the Digital Twin according to potential changes in infrastructure or
technology, thereby allowing them to effectively predict and analyze impacts on
system performance.

9 Unlocking Generative Urban Network Design Through

Digital Twins

Urban network design is entering a transformative era with the advent of genera-
tive modeling techniques. This is showcased in the development of a New Green-
field City network. By borrowing network components and the population’s activity
patterns from neighboring cities, planners can now simulate and visualize the
intricate dynamics of a new urban development even before it materializes. This
method leverages Mobiliti to create robust, predictive models that were previously
unattainable.
The generative design process for a New Greenfield City network is particularly
innovative because it synthesizes not just the physical infrastructure of the network
but also the underlying human activity that drives traffic patterns. This approach
includes borrowing and modifying both geometric elements and Origin–Destination
Time (ODT) data, providing a rich dataset that informs the design and anticipates
the needs of the urban area.
The effectiveness of this modeling approach is evident in the visualizations that
detail the origins and destinations of all regional trips including a New Greenfield
City network. These visualizations, Figs. 14 and 15, illustrate the density and spread
of trips in the region.
By adjusting the network elements—links added, removed, or modified—the City
Network Persona evolves. This persona reflects the changes in demand driven by
population activities and commercial vehicle flows, enabling a responsive design
process that adapts to predicted needs.
Initial model results highlight the existing Bay Area network with vehicle miles
traveled (VMT) before the integration of the New Greenfield City is shown in Fig. 16.
Figure 17 introduces the new city network in the upper right of the network, showing
how it integrates within the existing framework and the accompanying travel demand.
358 J. F. Macfarlane and I. Babur

Fig. 14 Origins of all trips taken by the population of a New Greenfield City network

Fig. 15 Destinations of all trips taken by the population of a New Greenfield City network

A zoomed-in view, Fig. 18, provides a detailed perspective of how the new network
segments interact with the existing network.
The final visualization, Fig. 19, demonstrates the regional impact of the New
Greenfield City on VMT. This includes not only the expected increase in VMT
across the entire Bay Area but also specific impacts on highway and county roads
in both the immediate and extended vicinities of Greenfield City. These areas show
significant changes, highlighted in the visual outputs, indicating a substantial increase
in traffic congestion and predicted pressure points. This analysis can be incorporated
as a feedback process for an automated generative design framework that can tune
the Greenfield City network and its associated demand to meet the design objectives
for the Greenfield City, while also addressing regional concerns associated with the
traffic impacts of the new travel demand.
Mobiliti: A Digital Twin for Regional Transportation Network Design … 359

Fig. 16 Daily vehicle count in the San Francisco Bay area before the addition of the New Greenfield
City network

Fig. 17 Daily vehicle count in the San Francisco Bay area after the addition of the New Greenfield
City network

The ability to perform such detailed and predictive analyses through Digital Twins
represents a significant leap forward in urban planning. Digital Twins enable plan-
ners to unlock insights that inform investment plans and policy decisions, facil-
itating a proactive approach to urban development. The generative design capa-
bilities facilitated by Mobiliti ensure that planners and stakeholders can explore
various scenarios and their consequences, enhancing understanding and preparation
for real-world implementations. This comprehensive approach not only optimizes the
360 J. F. Macfarlane and I. Babur

Fig. 18 Daily vehicle count in the New Greenfield City network after integration with the San
Francisco Bay area network

Fig. 19 Impact of the integration of the New Greenfield City network with the San Francisco Bay
area network in terms of daily vehicle count

network design but also ensures that new developments, such as the New Greenfield
City, are integrated seamlessly into the larger urban fabric, supporting sustainable
and efficient growth.
Mobiliti: A Digital Twin for Regional Transportation Network Design … 361

10 Future Research and Transportation Data Ecosystem

Development

The exploration of Digital Twins in transportation and urban planning is expanding.

The Mobiliti platform has prioritized vehicle traffic due to its significant impact on
health, safety, and quality of life in cities. The application of Digital Twins in urban
planning offers substantial potential for leveraging computational methods to tackle
complex real-world urban challenges. Traditionally, urban designs are developed by
human designers through extensive analysis, discussion, and iterations. However,
emerging computational models can streamline this process by optimizing designs
and generating multiple alternatives for designers to select from.
Computational AI methods, especially deep reinforcement learning (DRL) algo-
rithms, are being proposed to solve intricate challenges in urban science. Urban
design problems can be framed as sequential decision-making tasks with rewards
linked to various metrics, such as the creation of 15 min cities. With high-performance
computational support, the model can learn and gradually approach the optimal
solution, achieving strategies that surpass human capabilities.
When integrated with Mobiliti, using a multi-agent reinforcement learning
approach, a comprehensive urban planning, development, and management process
can be facilitated. This includes enhancing spatial designs by optimizing them with
respect to dynamic evolutionary outcomes like simulated traffic flow. Mobiliti is
applicable not only to individual cities but can also efficiently plan multiple cities
collectively, generating optimized plans that offer mutual benefits and yield more
efficient collective outcomes.
Several developmental areas could further unlock the full potential of Digital
Twins in transportation management:
. Digital Twin Taxonomies: There is a need to clarify the definition of Digital
Twins to resolve confusion in the data ecosystem. This could involve establishing
a clean definition or creating a taxonomy of Digital Twins with example instances,
enhancing data alignment, codifying physical system complexities, and defining
simulated models that capture primary components of the physical world but
acknowledge inherent randomness.
. Surrogate Models Trained by Simulation: Deep learning could capture complex
simulation aspects and offer model-based controller solutions.
. Data Access and Analytics: The transportation data ecosystem is rapidly
evolving. Enhancements in data availability and integration are crucial, alongside
developing common ontologies for network and land use to maximize value.
. Addressing Urban Complexity: As urbanization and the formation of mega-
regions continue, managing transportation in the context of city values will grow
in complexity. Accessibility to new computational approaches for urban planners
and broader ontologies covering domains like health, safety, and quality of life
will be essential for developing effective and equitable solutions and performance
indices that reflect the intricate interactions within city environments.
362 J. F. Macfarlane and I. Babur

Acknowledgements The Mobiliti Platform is the compilation of the work of a lot of excellent
scientists: Cy Chan, Srinath Ravulaparthy, and Max Bremer at Lawrence Berkeley Laboratory, Anu
Kuncheria, Dimitris Vlachogiannis, Colin Laurence, Anthony Patire, Rowland Awadagin Herbert-
Faulkner, and LuLu Lui at University of California Berkeley, Prasanna Balaprakash, and Tanwi
Mallick at Argonne National Laboratory. The work described was sponsored by the US Department
of Energy (DOE) Vehicle Technologies Office (VTO) under the Big Data Solutions for Mobility
Program, an initiative of the Energy Efficient Mobility Systems (EEMS) Program. The following
DOE Office of Energy Efficiency and Renewable Energy (EERE) managers played important roles
in establishing the project concept, advancing implementation, and providing ongoing guidance:
David Anderson and Prasad Gupte.

References

1. Chan C, Kuncheria A, Macfarlane J (2023) Simulating the impact of dynamic rerouting on

metropolitan-scale traffic systems. ACM Trans Model Comput Simul 33(1–2):1–29
2. Chan C, Wang B, Bachan J, Macfarlane J (2018) Mobiliti: scalable transportation simula-
tion using high-performance parallel computing. In: 2018 21st international conference on
intelligent transportation systems (ITSC), pp 634–641
3. Chan C, Kuncheria A, Zhao B, Cabannes T, Keimer A, Wang B, Bayen A, Macfarlane J (2021)
Quasi-dynamic traffic assignment using high performance computing. ArXiv:2104.12911 [Cs]
4. Faltesek A, Dakshinamoorthi B, Prabhala S, Thobani A, Kuncheria A, Macfarlane J (2021)
Urban traffic simulation: network and demand representation impacts on congestion metrics
5. Gasnet. [online]. Available: http://gasnet.lbl.gov
6. HERE Technologies—The world’s #1 location platform. en. url: https://wwwhere.com/(visited
on 11/11/2021)
7. Jefferson D (1985) Virtual time. ACM Trans Program Lang Syst 7, 3 (July 1985), 404–425
8. Koglin T, et al (2011) Measuring sustainability of transport in the city—development of an
indicator-set. https://doi.org/10.13140/RG.2.1.1451.0240
9. Kuncheria, A., Walker, J.L. and Macfarlane, J. (2023). Socially-aware evaluation framework
for transportation, Transportation Letters, pp. 1–18.
10. Kunz G (2010) Parallel discrete event simulation. In: Wehrle K, Güneş M, Gross J (eds)
Modeling and tools for network simulation. Springer, Berlin, Heidelberg. https://doi.org/10.
1007/978-3-642-12331-3_8
11. Macfarlane J (2021) Mobile device data analytics for next-generation traffic management
(Report)
12. Macfarlane J (2019) The transforming transportation ecosystem—a call to action
13. Macfarlane J (2019) When apps rule the road: the proliferation of navigation apps is causing
traffic chaos. It’s time to restore order. IEEE Spectr 56(10):22–27
14. Mallick T, Balaprakash P, Rask E, Macfarlane J (2021) Transfer learning with graph neural
networks for short-term highway traffic forecasting. In: 2020 25th international conference on
pattern recognition (ICPR), pp 10367–10374
15. Mallick T, Balaprakash P, Rask E, Macfarlane J (2020) Graph-partitioning-based diffusion
convolutional recurrent neural network for large-scale traffic forecasting. Transp Res Rec
2674(9):473–488
16. MATSim.org. [online]. Available: https://www.matsim.org/
17. Polaris. [online]. Available: https://polaris.es.anl.gov/
Digital Twin in the Battery Industry

Soumya Singh, Michael Oberle, Daniel Schel, Julian Grimm, Olga Meyer,
and Kai Peter Birke

Abstract The increase in global demand for lithium-ion battery cells has driven the
establishment of numerous manufacturing facilities. However, the complex design
and manufacturing of Li-ion battery cells pose several challenges, such as ensuring
high cell quality while maintaining high process stability and efficiency. During the
usage phase, factors such as operating conditions, and charging/discharging profiles
can significantly impact the performance and longevity of battery cells. Furthermore,
addressing recycling and proper disposal in the end-of-life phase presents substantial
challenges, given the risks associated with hazardous materials and the scarcity of
resources. Hence, effective management of these phases is essential for achieving
sustainable and efficient energy storage systems. Digital twins applied in various
industries have shown promising results in improving product lifecycle management
and smart manufacturing processes. However, it is important to scientifically test
and validate its effectiveness within the specific context of the battery industry. The
main question is whether digital twins can realistically address the growing chal-
lenges of the battery industry, such as degradation evaluation, usage optimization,
manufacturing inconsistencies, second-life application possibilities, etc. Without a
thorough understanding of the benefits of digital twins, it is difficult to claim that they
are the best solution for the battery industry’s challenges. This chapter explores the
potential applications and use cases of digital twins in the battery lifecycle, outlining
the requirements and concepts necessary for implementing digital twins effectively.
This approach ensures a comprehensive understanding of the system’s benefits
and limitations, thereby enabling stakeholders to make informed implementation
decisions.

Keywords Digital Twin · Battery lifecycle · Manufacturing · Li-ion battery · AAS

S. Singh (B) · M. Oberle · D. Schel · J. Grimm · O. Meyer · K. P. Birke

Fraunhofer Institute for Manufacturing Engineering and Automation IPA, Nobelstr. 12, 70569
Stuttgart, Germany
e-mail: [email protected]
K. P. Birke
Institute for Photovoltaics, Electrical Energy Storage Systems, University of Stuttgart,
Pfaffenwaldring 47, 70569 Stuttgart, Germany

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 363
M. Grieves and E. Hua (eds.), Digital Twins, Simulation, and the Metaverse, Simulation
Foundations, Methods and Applications, https://doi.org/10.1007/978-3-031-69107-2_15
364 S. Singh et al.

1 Introduction

Lithium-ion batteries (LiBs) are widely favored energy storage systems due to their
promising attributes such as high energy density, low self-discharge property and
long cycle life. The European Battery Alliance foresees a potential market worth
up to 250 billion euros for automotive batteries in Europe from 2025 onwards.
Presently, European production meets only 6% of global LiB demand, prompting
a political target to achieve 30% by 2030 [1]. To support this goal, the Federal
Ministry for Economic Affairs and Energy has allocated nearly three billion euros
for battery cell production, attracting proposals from over 40 German companies [2].
This funding prioritizes sustainability and digitalization, aiming for efficient material
usage, energy conservation, and innovative recycling approaches. In order to meet
such requirements and the rising global demand, the challenges across the battery
lifecycle need to be addressed through cutting-edge technologies [3].
Driven by the rising market demand for battery electric vehicles (BEVs), consumer
electronics, and battery-driven grid storage, stakeholders across the battery value
chain encounter numerous challenges throughout the battery’s design, manufac-
turing, usage, and end-of-life (EoL) phases. This section will subsequently explore
the challenges within the different lifecycle stages.
During battery design, challenges can arise from market demands, such as
customers along the supply chain, including original equipment manufacturers
(OEMs), or end-users. Regarding BEVs, users demand high energy density, a long
driving range, long calendar and cycle life, fast and efficient charging, safety, and
low costs [4]. On the other hand, OEMs face design limitations because they aim
for application-specific batteries. The variety of applications potentially influences
the challenges in the design stage [5]. Furthermore, the battery market is dynamic
at the moment, with heterogeneous manufacturers, different packs/cell designs and
formats, electrode designs and formats, different active and inactive material combi-
nations with various levels of readiness, as well as promising future technologies [6,
7]. Bringing all this together, battery design has multiple and fast-changing objec-
tives, conditions, and circumstances, which, as a result, leads to tradeoffs and batteries
that cannot meet all aspects.
The demands within the battery market generate technical challenges for battery
design and development. It is crucial to reach good performance measures with a
specific battery design. Reaching a high energy density is said to be the overarching
goal [8]. Besides this, safety, thermal stability and temperature ranges, and durability
are further influencing factors.
The aim of achieving a high energy density and fast charging rates, results in
challenging cooling efforts on the battery system for dissipating and supplying heat.
Furthermore, increasing the size and thickness of individual cells to meet the target
values changes the electrical and thermal conditions at a cell level in a negative
manner [8]. Thus, reducing the amount of inactive material in either a battery cell or
Digital Twin in the Battery Industry 365

a battery system may lead to undesired conditions. In contrast, the system’s thermal
management requirements can be reduced by increasing the cell’s thermal operation
window [5]. In addition to thermal requirements, there is also a tradeoff between
general safety and energy density. In a battery system, there are electrical, thermal,
and mechanical safety protection measures on the cell, module, pack, and vehicle
levels [5]. This results in an effective safety system but ends up in a lower energy
density. Furthermore, battery durability is a critical parameter due to warranty issues
of the battery manufacturer and OEMs. Therefore, in the design phase, selecting
suitable tests and confirming test results for possible future degradation of the battery
during usage under operating conditions is challenging [5].
Aside from the design challenges due to technical requirements, there are also
issues arising from scarce and expensive raw materials. Since mining and refining
operations are concentrated in a few countries, future resource supply could become
challenging [9]. Efforts are underway to reduce reliance on scarce resources like
cobalt and nickel in battery production [5]. With large-scale battery facilities being
developed, it is crucial to understand how changes in battery design might affect
manufacturing processes and supply chains. Overall, battery design encounters
challenges due to tradeoffs in target values and evolving conditions [7].
In battery manufacturing, the main objectives are cost savings, quality improve-
ment and fostering sustainability. However, achieving these objectives present chal-
lenges such as scaling up production, enhancing productivity, minimizing scrap, and
optimizing resource and energy efficiency [10]. These challenges can be addressed
at different levels, such as manufacturing operations or process levels.
As elaborated in [10], certain manufacturing processes have prolonged processing
times, consequently reducing the yield rate. Notably, the electrolyte filling process
stands out as one of the most time-consuming steps. With an increase in the cell’s size,
the reduction of the electrolyte filling process time becomes even more challenging.
Another example related to large cell formats with higher energy density involves
the necessity of adjusting their tab designs according to the specific format, resulting
in the need for greater positioning accuracy and posing handling challenges [8]. Last
but not least, modification in cell designs, as well as the integration of new battery
designs into existing production lines, can be very challenging because they require
significant adjustments to manufacturing processes, machinery, and workflow, often
leading to disruptions, increased costs, and potential delays in production schedules
[7].
During battery usage, challenges may arise due to user expectations, particularly
concerning the total costs of ownership (TCO), including purchasing costs, electricity
costs, and maintenance costs. To mitigate these expenses, efforts must be made across
the design, manufacturing, and EoL stages, involving process improvements, material
innovations, and implementing circular economy principles [5].
A major challenge for LiBs is that different operating conditions and load profiles
create individual life cycles that greatly affect the aging of a cell. Consequently,
the aging of each LiB cell within a battery pack differs from one another, leading
to internal inconsistency between the battery cells within one module or pack
[11]. Lastly, challenges such as monitoring performance, ensuring safety, managing
366 S. Singh et al.

thermal conditions, access to charging infrastructure, and accurately predicting

battery life are crucial for maintaining efficient battery operation and extending
longevity.
Generally, the EoL of batteries involves processes such as recollection, sorting,
dismantling, and material recovery. Currently, the return rate of used batteries is
low [11]. The methods used for processing batteries at EoL can be as unique as the
design of each individual cell. Overall, research in the field of EoL batteries is still
ongoing and there are mainly no standardized processes, making it very challenging
[12, 13]. It is unclear which batteries should be processed for which circular value
creation solution, e.g., reuse, remanufacturing, repurposing, repair, and recycling.
Safety is a major issue for all these solutions. Once a used battery is returned, the
battery’s and cell’s condition is unclear, but they are still electrochemically active and
contain chemically stored energy and flammable materials. During dismantling, the
batteries are treated mechanically and are thus exposed to further potential damages.
For reuse, batteries should generally be available at a high remaining capacity and
low cost [12].
Currently, second-life applications are in development and testing, and there are
not yet proven technologies or processing standards. Additionally, ongoing research
on recycling [5, 13] faces the challenge of separating tightly coupled materials within
battery systems (dismantling) and battery cells (recycling). Standardized procedures
for dismantling and recycling are currently lacking, and there are various process
routes for handling different battery chemistries [13].
Besides challenges related to standardization, safety, and future market develop-
ments, new regulatory regulations also influence activities in the EoL of batteries.
This includes, for example, the EU Battery Regulation called Battery Passport, which
mandates the use of digital passports for EVs and industrial batteries over 2 kWh
sold into the EU market starting from February 2027. These passports must contain
essential information accessible to the general public, regulatory bodies, and battery
service providers. The Battery Passport aims to enhance sustainability and circularity
in battery value chains by promoting transparency, facilitating efficient recycling, and
encouraging sustainable practices among manufacturers.
In summary, the battery life cycle presents significant challenges in design, manu-
facturing, usage, and EoL stages. This chapter explored the applications of the
concept of Digital Twin (DT) as a promising solution to address these challenges.
By aligning DT use cases with industry challenges, the potential benefits of a DT
implementation become more evident. The discussion delves into specific use cases
within different lifecycle stages, supported by an implementation framework and a
case study example.

1.1 Digital Twin Definition

In 2002, the basic concept of DTs was first presented as “The Conceptual Ideal of
PLM” and later in 2005 as the Mirrored Spaces Model by Dr. Michael Grieves. This
Digital Twin in the Battery Industry 367

was the first time that the idea of virtual representation of real objects was applied
in the context of industrial production. Dr. Grieves, who was the Co-director of a
research center at the University of Michigan at the time, is considered a pioneer
who put the concept of DT into practice [14, 15].
DT technology enables the recreation of real objects or systems in a virtual envi-
ronment. Data is collected from sensors and other sources to create a replica of
the real object. The DT can then be used for various purposes, such as simulating
operating procedures, optimizing processes, or predicting malfunctions.
The term “Digital Twin” was introduced by John Vickers from NASA in 2010
[15, 16]. It serves as a metaphor illustrating the idea that the DT is a virtual copy
of a real object. Similar to real-life twins, the real object and the DT share certain
characteristics and features. Since then, the number of publications on the subject
has increased enormously [17].
Jones et al. in [18] recorded a sharp increase in publications on the topic of
DTs as early as 2014. Despite the significant increase in the interest in the DTs in
recent years, research is still at a very early stage, with initial practical applications
remaining tentative and not fully developed. As evidenced by various publications
on DTs in [18, 19], there was already a pressing need to consolidate research and
industry in the early 2020s to establish a common understanding of the topic and
ensure that future research efforts are based on solid foundations for each specific
area.
The term “Digital Twin” has become ambiguous due to the abundance of scientific
publications and industrial application scenarios. Consequently, organizations have
been formed at the national and international levels with the goal of promoting
and disseminating the concept of DT. However, there exists a range of definitions,
graphic models, and complementary concepts. Nevertheless, the lowest common
denominator in the context of DT remains the Mirrored Spaces Model by Grieves
[20, 21].
The conceptual process is significantly delayed by the need to define terminology
and concepts for the DT. Consequently, numerous definitions have been broadly used
in both industry and academia for some time now, posing a major challenge for users
to find the right definition and base their work on useful practical specifications. The
lack of consistency and a number of non-consensual solutions, based on a variety
of different definitions and frameworks can potentially risk the benefits that the DT
was originally intended to provide [18].
For several years now, standardization has been actively pursuing a coordinated
approach to the topic of DTs. The forerunner in this area is the Joint Technical
Committee of ISO and IEC (JTC 1), established to develop global standards for infor-
mation and communications technology (ICT) for business and consumer applica-
tions. Within JTC 1, there is a special subcommittee, SC 41, dedicated to standards for
the Internet of Things. Since the beginning of 2021, this subcommittee has expanded
its scope and terms of reference by establishing a new working group (WG 6) dedi-
cated to DT applications across multiple sectors, including smart manufacturing,
utilities, smart grids, smart buildings, and smart cities [22].
368 S. Singh et al.

Unfortunately, due to the high interest and participation of numerous experts

from many countries at the international level, coordination has slowed down and
the full description of the concept for different areas has become a lengthy process.
Nevertheless, the first basic standards have already been published in initial drafts or
are officially due for publication, which may lead to initial successes in addressing
definitional confusion.
At this stage, we are rely on the definition of the Industry 4.0 platform, which
forms the basis for the rest of this chapter:
A digital twin is a digital representation, information in the digital information world that
represents characteristics and behaviors of an entity, sufficient to meet the requirements [23].

It took some time to agree on the definition and formulation of the basic terms,
however, the situation is different in practical research. Here, developers started
a bit earlier and came up with the first solutions. A glimmer of hope is coming
from the field of manufacturing. Due to the advancing digitization in the industrial
environment, the development of concepts and closer practical solutions for the DT
is being tackled more quickly than in other areas.
One of these areas is battery cell production. In the following sections, we will
take a closer look at the product, the battery cell, or rather the life cycle of battery
cells and the possible applications for DTs in this context.

1.2 Digital Twin Concept

As previously mentioned, Grieves’ Mirrored Spaces Model is the lowest common

conceptual denominator in the context of DT. It describes how the real space is
connected to one or more virtual spaces (VS) through a linking mechanism based on
data, information, and processes. The various virtual spaces show different represen-
tations of the same physical object, facilitating the exploration of alternative designs,
among other purposes. Grieves emphasizes that the flow of data and information in
its original state is regularly processed and forwarded by humans. To realize its full
potential, a “robust, accurate, and timely [14]” link is essential. While numerous other
concepts exist, some build on this framework while others offer different perspectives
on DT [21].
However, such a definition alone is not sufficient to describe or classify the various
application scenarios of DT. Grieves addresses this by introducing three types of DT
along the product lifecycle.
The Digital Twin Prototype (DTP) includes all potential products that can be
created. It represents the product and its various versions as it progresses step by
step, from the initial idea to the production of the first manufactured item.
The Digital Twin Instance (DTI) is created at the start of production. Each product
that is created can have a DTI, which represents the collection of different types
of information for the respective product instance, such as serial numbers, service
Digital Twin in the Battery Industry 369

information, part measurements, assembly process information, etc. Much of the

information is fed from the product’s DTP, which is linked to the DTIs.
The Digital Twin Aggregate (DTA) is the aggregate of all manufactured products.
It represents the collection and aggregation of data from the product population and
thus forms the summary of all DTIs, always with the aim of creating value [15].
These DT types are of a basic nature, meaning the complexity regarding the product
lifecycle, the entities involved in it, and their relationships to each other are greatly
reduced, allowing only a rudimentary classification of the application scenarios of
DT.
Another interesting concept that supports the DT classification is the Digital Twin
Structure Model (DTSM), developed despite the lack of a coherent DT definition and
based on the information and knowledge about DT available by 2020. This model
incorporates various key elements, including an asset lifecycle perspective integrated
with the Product, Process, and Resource (PPR) model. It facilitates the structuring,
categorization, and comparison of DT use cases [20].
We will only discuss the most important aspects of the DTSM and relate them to
the Grieves DT types. Fig. 1 provides a coherent description scheme for application
scenarios of DT. The DTSM essentially consists of three dimensions: Lifecycle, PPR,
and Cyber-Physical.
1. The first dimension, “Lifecycle”, consists of two phases: type and instance. These
phases are identical to the Grieves DT types. The DTP is relevant in the type
phase, while in the instance phase, the DTI and DTA represent multiple DTIs.
2. The second dimension, the “PPR” model, describes the distinction between
product, process, and resource and is widely assumed in production and manu-
facturing (see [24, 25]). It is a helpful model to describe the relationships and
entities involved in the product creation process in sufficient detail. According
to the DTSM, products, processes, and resource have their own lifecycle and can
therefore be represented by their own DTs. We make use of this in the description
of the application scenarios, e.g., to differentiate between the DTI of a product
and a resource that generates it.
3. The third dimension, “Cyber-Physical”, describes the relationship between the
physical world and the virtual world in the context of the DT. An asset (i.e.,
product, process, resource) exists in the physical world and is represented by the
cyber part (i.e., DT) in the virtual world. A further distinction is made in the
cyber part between two levels: Descriptive and Executive. These describe the
static (i.e., descriptive) and dynamic (i.e., executive) aspects of the DT and are
therefore very well suited to classify the behavior of a DT.
This shows that the two concepts, the DTSM and the Grieves DT types, can
complement each other. The combination of these two concepts serves as a foundation
for exploring DT in the battery industry, as outlined in this chapter. To this end, the
applicability of DT in the battery industry is described based on identified use cases
along the battery life cycle with the aid of both the concepts.
370 S. Singh et al.

Fig. 1 Classification of DTs based on [20] and [15]

1.3 Battery Lifecycle

Product Lifecycle Management (PLM) is the business activity of managing a product

throughout its lifecycle, from the initial idea to its retirement and disposal [26]. The
concept of DTs originally emerged to address PLM but has since been extended
to include battery lifecycle management [27]. There isn’t a universally recognized
PLM standard dedicated solely to the battery lifecycle. Instead, PLM standards are
typically used to manage the lifecycle of products and systems overall, which may
include LiBs and their components. The concept of product lifecycle includes the
following major stages: conceptualization, organization of resources, design, anal-
ysis, simulation, and verification, manufacturing and distribution, in-use performance
and customer feedback, and eventual disposal [28]. In this context, a battery lifecycle
can be described according to the nodes and loops represented in Fig. 2.
LiB lifecycle primarily starts with raw material extraction and its supply to elec-
trode manufacturers. Battery systems can be divided into three primary components
and their respective production stages: electrode manufacturing, cell manufacturing,
and module/pack manufacturing. According to PLM standards, each of these compo-
nents can undergo individual management stages such as product ideation and orga-
nization design, conceptual design, detailed design, design evaluation, production
planning, production, distribution, usage/service life, second life, and EoL. The
service life of a battery system is connected with the usage of the entire battery
module or pack within its respective application. Lastly, the EoL stage can involve
direct recycling, remanufacturing, reuse, or repurposing at both the cell and system
levels.
DTs applied in various industries have shown promising results in improving
PLM and smart manufacturing processes. However, it is important to scientifically
Digital Twin in the Battery Industry 371

test and validate their effectiveness in the specific context of the battery industry.
Without a thorough understanding and substantial efforts in quantifying the benefits
of DT, it is challenging to assert that it is a suitable solution to the battery industry’s
challenges [29].
However, DT implementations present their own set of challenges, such as
ensuring data accuracy and integrity, managing the complexity of integrating various
data sources, addressing cybersecurity risks, and ensuring scalability and interoper-
ability. Additionally, there may be challenges associated with the cost of implemen-
tation, organizational resistance to change, and the need for specialized expertise in
areas such as data analytics and machine learning.
This brings us to the research questions addressed in this chapter. Firstly, what are
the potential applications of DTs across various stages of the battery lifecycle? We
identify specific use cases where strategic implementation of DTs, as hypothesized,
could alleviate the challenges currently faced. Secondly, how can the development
and implementation of a DT be optimized for a battery-specific industrial context?
By delving into both the theoretical and practical dimensions of these inquiries, we
aim to contribute insights into leveraging DTs for the battery industry.

2 Potential Use Cases of Digital Twins in the Battery

Lifecycle

While it is essential to identify use cases and understand the potential business advan-
tages and functional benefits of DT applications, accurately quantifying these benefits
remains challenging due to the current technological landscape and the ambiguity
surrounding DT classification. Existing literature [3, 27, 30] highlights the theoret-
ical analyses of DT applications in the battery lifecycle; however, translating these
theoretical insights into tangible business advantages necessitates actual implemen-
tation. Recent studies emphasize the need to move beyond theoretical discussions
to quantitatively assess the real-world impacts of DT utilization. In this context,
Fig. 3 summarizes the potential challenges within the battery lifecycle that could be
mitigated through the practical implementation of DT.
In the following sub-sections, we explore the potential use cases of DTs within
each lifecycle phase to address these challenges.

2.1 Digital Twin for Battery Design

2.1.1 Objective

With the continuous evolution of battery materials, the pursuit of an optimal battery
remains ongoing, requiring the introduction of new designs and technologies at both
372 S. Singh et al.

cell and pack levels. This process is both costly and time-intensive. A crucial use case
of DTs is in providing decision-support for battery system design by leveraging real-
world data to reflect the battery’s behavior and performance in various operational
and environmental conditions.

2.1.2 Overview

The trial and error approach in developing new battery designs often results in signif-
icant consumption and wastage of critical raw materials, such as nickel and cobalt.
The motivation behind using a DT during the design phase is to provide stakeholders
with a decision-supporting tool for streamlining the design processes, thereby mini-
mizing material waste and optimizing resource utilization. It aims to facilitate the
exchange of data and information across different lifecycle stages as well as the
prediction of battery behavior and performances using data from various operating
conditions and data from previous product generations, hence, aiding in the design
of more durable and efficient battery system. Additionally, the operational data of
the battery can be used as feedback to adjust the design and optimization processes.
During the standard design process of LiB systems, several key features need
consideration. For cell design, aspects such as nominal capacity, voltage, internal
resistance, energy and power requirements, cell format (cylindrical, prismatic,
pouch), and electrode material (e.g., lithium cobalt oxide, lithium iron phosphate)
play a crucial role. Safety features, including prevention of overcharging, discharging,
and thermal runaway, along with cooling systems and electronic control through
Battery Management System (BMS), are vital considerations [31].
Recent studies [32, 33] have proposed various DT frameworks for smart product
design. Applying DT in the product design stages allows for providing a quantita-
tive design tool for efficient and optimal design decisions using data from previous
product generations [34], as well as operational data from the usage phase. In the
design process, battery models, such as electrochemical-thermal models or other
simulation-based and data-driven methods, play a crucial role in scenario simula-
tion. These tools are instrumental for enhancing battery design, reducing long design
iterations, and expediting the time-to-market.

2.1.3 Implementation Framework

This use case falls under the category of a DTP of the product. According to Fig. 1,
there is a physical component and a cyber-component of the product DTP. Aligning
the use cases according to the classification system in Fig. 1 helps establish the
system boundary for implementation. This process of defining the system boundary
will be applied to all subsequent use cases outlined in the following sections.
The battery DTP primarily comprises a combination of data, models, and algo-
rithms already employed for battery simulations. In the design phase, the battery
Digital Twin in the Battery Industry 373

Fig. 2 Battery lifecycle

Fig. 3 Challenges across the battery lifecycle, potentially addressed by DTs

DTP can determine the impact of various parameters on system behavior, identify
problems within the system, and simulate deterioration deterrence options [35].
Figure 4 depicts the DTP for battery design. On the left side of the figure, each life
cycle phase is associated with the types of DTs based on their definitions (in Sect. 1.2),
namely DTP, DTI, and DTA. The battery DTP streamlines the design processes by
utilizing the design data, experimental data, real-time and environmental-impacted
data to feed the battery models. Various modeling approaches are available for simu-
lating the battery states (such as state of health, state of charge, etc.), performance,
aging and safety of LiBs. These include electrochemical models (like pseudo-2D-
P2D model, single particle model SPM, etc.), electrical models (equivalent circuit
model ECM), thermal models, mechanical models, empirical models, data-driven
approaches, and hybrid approaches. These models are coupled with algorithms, listed
in Fig. 4, to streamline the scenario simulations.
The DTP serves as a virtual prototype, simulating design scenarios and facilitating
efficient testing and redesigning. This reduces costs by identifying flaws before actual
production. By harnessing the capabilities of the DTP, battery designers can create
374 S. Singh et al.

Fig. 4 DT for battery design

dynamic simulation scenarios, optimize auxiliary systems, and accurately predict the
performance of physical products, enhancing design efficiency and product quality.
Let’s consider a scenario where an existing battery pack is currently utilized in
EVs. Now, there arises a new imperative for EV manufacturers: the design of a novel
battery pack featuring a new cell chemistry that promises higher energy density.
Notably, all other aspects of the pack will remain unaltered. In light of this, we
explore the use case of a battery DTP within this context. At the model level, changes
in cell chemistry affect a wide range of electrochemical parameters, such as particle
radius, electrode thickness, conductivity, and diffusivity of the active materials. These
affected parameters can be provided by the material supplier, measured through test
experiments, or identified through various parameter identification algorithms. The
existing pack design also has DTI and DTA from the manufacturing, usage, and EoL
phases. Using the information on the new design requirements and the DTI/DTAs of
the previous product version, a DTP of the new product prototype is developed.
The DTP of the new design offers a transformative solution for battery designers
by integrating data from various lifecycle phases, facilitating simulation, testing,
and optimization of battery performance in a virtual environment. Consequently,
before initiating the entire value chain, the compatibility of the new design at both
the module/pack and system levels can be tested virtually. When designing or opti-
mizing the current battery design to align with operational needs, the real-time data
provided by the DTIs offer insights into the battery pack’s operating cycle, working
environment, and other critical information not easily obtained through traditional
laboratory experiments. As a result, the DTP promotes battery design, development,
and research, ushering in a more efficient and informed approach to advancing battery
technology.
Digital Twin in the Battery Industry 375

2.1.4 Challenges

. The models used for simulation the design scenarios are the core element of the
DTP. Ideally, the battery model is a high-fidelity multiscale that can incorporate
both temporal and spatial dynamic features. One of the challenges is to quantify
and manage the uncertainty inherent in these scenario simulations and subse-
quence choices based on them. Furthermore, the verification of such models is
complex and computationally intensive.
. The challenge associated with the effective integration of data from various life
cycle phases and technical domains during the implementation of a DTP for
battery design optimization lies in ensuring seamless communication and compat-
ibility between disparate datasets. Achieving this integration requires overcoming
technical barriers, standardizing data formats, addressing interoperability issues,
and ensuring data consistency and accuracy across different sources.
. The aggregation of multiple datasets used to feed the models is associated with
multidisciplinary challenges concerning the sharing methods, privacy, security,
and governance systems.

Overcoming these challenges requires robust data integration mechanisms,

advanced modeling techniques, and seamless communication between the battery
systems and the DTP. The concept of DTI plays an integral role here, potentially
serving as a bridge between the diverse data sources and the DTP, facilitating real-
time data exchange and enhancing the accuracy and reliability of the simulation and
optimization processes.

2.2 Digital Twin During Battery Manufacturing

2.2.1 Objective

Batteries are required to adhere to strict performance and quality standards. However,
manufacturing processes include non-deterministic factors like material inconsis-
tencies, environmental conditions, human error, and equipment wear. These factors
contribute to variability in the outcomes of the manufacturing process. The implemen-
tation of DTs in battery manufacturing is aimed at mitigating these inconsistencies in
order to minimize scrap, increase throughput, and enhance reliability of the battery
cell during use.

2.2.2 Overview

To achieve the mentioned objectives, different types of DTs are required, each serving
specific purposes. If the aim is to reduce scrap and ensure cell reliability, the DT of the
process becomes paramount. Conversely, if the objective is to increase throughput,
376 S. Singh et al.

Fig. 5 DT during battery manufacturing

then the DT of the resource is key. Additionally, the DTI of the product also plays a
crucial role in supporting both objectives.
While the DTP of the product is critical during product design and provides the
baseline structure and data for the DTI, the DTI of the product becomes key during
the manufacturing life cycle. There exists an essential link between the DTI of the
product and the DTI of the process and resource. On one hand, it collects data
throughout the entire manufacturing process, while on the other, it identifies if the
asset is within specifications. DTAs of the product enable statistical analysis and
machine learning approaches to derive models and rules that can be leveraged for
the creation of DTP of the processes.
In a use case aimed at ensuring cell quality by maintaining manufacturing process
outcomes consistently within specifications, the DTI of the process utilizes informa-
tion from the product DTI regarding process results and product specifications for
its calculations. For example, this can allow for the dynamic calculation of a recipe
for adjusting the dryer during coating based on the viscosity of the slurry. Similarly,
the DTI of the product supports the DTI of resources in the battery manufacturing
process. Information on the result of each process step is essential information for
the DTI model in terms of calculating the wear level. Based on this information, the
DTI can determine when maintenance procedures need to take place to minimize
resource downtime (Fig. 5).

2.2.3 Implementation Framework

Considering the DTSM, both the DTI of the product and the DTI of the process and
resource have descriptive and executive components. Figure 6 shows a framework for
Digital Twin in the Battery Industry 377

Fig. 6 DT during operation

DTs in battery manufacturing for PPR and illustrates their interrelations, including
the primary data required.
For the DTI of the product, quality data is collected at each transformative stage
of the manufacturing process. This data is continuously enriched from the mixing of
slurry to the manufacturing of electrodes, cell assembly, and finally, cell formation.
The collection of quality data is particularly crucial during manufacturing, not only
for compliance with customer standards but also for internal processes.
The product’s DTI includes an executive component for evaluating the current
state of its quality, employing either statistical or machine learning models. Alterna-
tively, it can utilize the DTP and its models developed during battery design. Quality
data collected up to the current work in progress is utilized for this purpose. For
example, in slurry mixing, quality data encompasses viscosity and homogeneity.
Quality models can assess whether these values fall within acceptable parameters
and predict their impact on future manufacturing steps. Typically, these models are
constructed using machine learning [36, 37].
On the manufacturing shop floor, each process has a separate DTI due to its
distinct nature. For example, the mixing and coating processes are continual, whereas
assembly and formation are discrete. It is evident that each manufacturing step can
differ significantly from another, even without considering the specific techniques
involved. The process DTI comprises descriptive and executive components. The
descriptive parts include all parameters (e.g., time or temperature for mixing) that
define the process. The executive part models the process behavior with the aim of
providing actionable information for process control. Typically, this involves a recipe
that includes process parameters for the current run of the process.
Information of the product DTI is crucial for the process DTIs. On one hand,
it facilitates adapting the process to the current product by retrieving the current
characteristics achieved by preceding processes. On the other hand, it can provide
indications of effectiveness by retrieving the outcomes of subsequent processes to
adapt for future process runs [38]. These dynamic short-term adjustments enabled
by the DTAs of the product allow for more fundamental process adjustments by
conducting deeper analyses on hundreds, if not millions, of occurrences [39].
The resource DTI functions similarly to the process DTI in interactions and
capabilities but focuses on machine data rather than process data. For instance,
in a battery manufacturing sub-process like tab welding, executed using different
378 S. Singh et al.

machines such as laser and resistance welders, the process DTIs control the process,
while resource DTIs handle energy consumption and machine behavior. Models
derived from resource DTIs optimize machine-related factors such as maintenance
scheduling or cycle times, employing techniques such as discrete event simulation
[40] or machine learning models [41].

2.2.4 Challenges

. As described in the previous paragraphs, the interplay of the PPR DTIs plays
a crucial role. However, besides communicating with each other, this requires
gathering data from various systems, such as ERP, MES, machines, and tools.
Integrating this data from diverse sources and ensuring its accuracy and timeliness
is a significant challenge. This is exacerbated by the vast amount of data that
needs to be aggregated. Considering that several hundred thousand battery cells
are produced each day, the data volume that needs to be processed and stored is
immense. Hence, careful planning of the system architecture is required.
. In addition, the creation and maintenance of the DTIs for process and resource,
particularly the executive models, can be challenging. For both, accurately
modeling the dynamic behavior of the real asset is essential. This in particularly
challenging in battery manufacturing processes, which are complex to model. A
prime example is the coating process. While the fundamental steps required for
coating are well understood, modeling becomes difficult due to numerous influ-
encing factors (e.g., slot control, dryer control) that affect the behavior of the
model. Furthermore, the relationship between processing conditions (e.g., slurry
composition, coating speed, drying temperature) and the resulting microstructure
is an area of active research.
Addressing these challenges requires strategic integration and a robust system
architecture. One potential approach is to implement standardized description
formats and interfaces that enable uniform models and access to diverse data sources.
This approach could also facilitate abstracting physical data storage, separating repre-
sentation from big data high architecture design. Furthermore, the creation and
maintenance of executive models could be streamlined by integrating simulation
models and machine techniques. This integration can help mitigate situations where
a lack of data impedes machine learning efforts, or validate data-driven models with
simulations.
Digital Twin in the Battery Industry 379

2.3 Digital Twin During Battery Usage

2.3.1 Objective

The objective of a DT during the battery usage phase is to predict and investigate the
degradation mechanisms and potential failures, thereby optimizing performance and
extending the service life. Through the integration of battery operation data with the
models, a DT enables proactive management of battery health and performance.

2.3.2 Overview

The ever-increasing demand for efficient energy storage solutions necessitates a

deeper understanding of battery behavior throughout its lifecycle. This involves
ensuring safety during operation and unlocking its full potential in terms of perfor-
mance and service life. A battery DTI can potentially serve as a powerful tool, to
comprehensively investigate degradation mechanisms and enhance the performance
and longevity of batteries during usage.
The primary goal of implementing a DTI in this context is twofold. Firstly, it aims
to advance our capability to foresee and comprehend the degradation mechanisms
and potential failures that batteries may undergo during usage. Secondly, the DTI
serves as a predictive and optimization tool, enabling a proactive approach to enhance
battery behavior. By seamlessly integrating real-time data and simulations, the DTI
facilitates not only a better understanding of battery dynamics but also empowers
strategies for optimizing their overall performance and extending longevity. This
pursuit is pivotal in ensuring the reliability and effectiveness of battery systems in
various applications, from portable electronics to electric vehicles and renewable
energy storage.
The potential functionalities of a battery DTI during usage and to some extent, also
for EoL are identified in [29], namely: product design and optimization; behavioral
integration in other life cycle phases; Prediction of Remaining-Useful-Life (RUL)
prediction; evaluation of battery ageing indicators; charging optimization; adaptive
control of thermal management; fault diagnosis. In a technical context, the results
that such a DTI can offer are as follows:
1. Estimation and prediction of State of Charge (SOC) and State of Health (SOH)
2. Prediction of RUL of the battery cell/module
3. Prediction of voltage response
4. Measures for usage optimization
Holistic approaches for implementing DTs for the battery operation phase have
been sparingly explored in the literature except in [29, 42]. In general, most of
these approaches are a combination of the following components: physics-based
model, data-driven model or a hybrid model; model parameterization approach; a data
management module; model evolution approach; and result calculation algorithm.
380 S. Singh et al.

Together, these components act as a backbone intelligence that can be uploaded on

the BMS.

2.3.3 Implementation Framework

This use case falls under the category of DTI of a product, specifically the battery
system. According to Fig. 1, there are both physical and cyber components of the
product DTI. In the context of this use case, a DTI represents a collection of data
generated when the battery system is under operation in a certain application, such
as the data from the BMS, service information, serial number, etc.
The DTI of the individual cells, modules, and battery systems are already created
during the manufacturing phase. Battery systems are complex products composed
of multiple modules and a significant number of battery cells. The DTI of a battery
system is composed of the DTAs of the cells and the DTAs of the modules. During
the usage phase, the DTI of the battery system is connected to the specific physical
twin and remains connected throughout the entire life of that physical twin (i.e., the
battery system).
Figure 6 represents a battery system DTI. The behavioral profile of the battery
is recorded as operational data during its actual usage. This typically includes data
acquired from on-board BMSs, consisting of information from current, voltage, and
temperature sensors. In most cases, BMS data is stored locally. However, recent
research efforts have focused on mapping real-time battery data from BMS to cloud-
enabled systems. This approach can potentially minimize local computational needs
and aggregate large datasets to improve the performance of ML-based algorithms.
The DTI can continue to utilize the models and algorithms in the DTP created
during the design phase. By feeding operational data into the DTP developed during
the design phase, the battery model needs to evolve to align with battery usage
profiles. This ability for model evolution is a vital component of the algorithm,
providing the battery DTI with essential “intelligence” during the usage phase.
The exploration of model evolution represents an intriguing research gap within
battery modeling and is the focus of extensive investigation. This concept is further
elaborated upon and applied in a case study presented in Sect. 3.
The DTI’s result for this use case involves degradation assessment in terms of SOH
estimation and RUL forecast, along with optimization measures based on predicting
voltage response. These capabilities enable the derivation of maintenance measures
during servicing, facilitating efficient troubleshooting, improving customer satisfac-
tion, and providing advanced monitoring capabilities. While individual users may
not delve deeply into battery degradation mechanisms, organizations utilizing battery
systems for applications such as EVs, smart grids, or industrial power require a
comprehensive understanding of how the usage patterns influence system aging.
These users benefit from the results of a battery DTI in terms of optimizing charging
patterns, accurate range estimation, estimating system reliability, and responsible
energy management.
Digital Twin in the Battery Industry 381

2.3.4 Challenges

. The DTI of a battery system ultimately constitutes a DTA of multiple modules.

Such an architecture can be challenging to develop and maintain, particularly in
the absence of standardized reference DTIs at the cell and module levels.
. Development and verification of a LiB model that efficiently emulates the real-
time behavior of the system is complex and requires massive amounts of data.
Moreover, the method of model evolution is not well-established and still needs
further research. Additionally, accurately estimating the changing parameters for
the battery model is challenging due to inherent cell-to-cell variations arising from
subtle manufacturing differences, resulting in non-linear degradation over time.
. Integrating operational data into the model can be challenging. While cloud
services have been used with onboard BMSs, standardizing data integration
methods for battery DTs is crucial, despite requiring initial investment and
implementation time.

To address these challenges, standardizing reference DTIs at both the cell and
module levels is crucial. This will simplify the development and maintenance of
the architecture. Additionally, progress in Li-ion battery modeling techniques, along
with the establishment of standardized data integration methods, and further research
in model evolution methods are essential steps. These endeavors will enhance archi-
tecture development, improve model accuracy, streamline data integration, and refine
battery models, ultimately optimizing system performance and fostering innovation
in the energy storage sector.

2.4 Digital Twin for End-of-Life Applications

2.4.1 Objective

The objective of a DT at battery EoL is to enable informed decision-making for

optimal second-life applications and to ensure effective and safe remanufacturing,
recycling, and repurposing of batteries. It serves as a bridge, filling the informational
gap that exists at EoL concerning other life cycle phases.

2.4.2 Overview

With the increase in demand for LiBs, prolonging the overall utility of the batteries
is becoming increasingly important. Due to the presence of critical materials, such
as cobalt and nickel, and the high manufacturing costs of battery packs, there is
economic, political, and environmental interest in utilizing existing battery system
in second-life applications. Currently, there is insufficient information available for
choosing a second-life application. Examples of second-life applications include
382 S. Singh et al.

power grid storage systems, automated guided vehicles like forklifts, and home
storage systems for storing photovoltaic electricity temporarily. In such applications,
the battery system experiences lower demands in terms of external temperatures,
charge/discharge currents, and SOH.
DTs provide an approach to tackle this challenge by creating a digital representa-
tion of the battery´s past life, including design details, manufacturing details, usage
history, and environmental interactions. This provides a comprehensive insight into
both the structural and functional aspects of the batteries. Such information is crucial
for designing, planning, and executing the processes of remanufacturing, recycling,
and repurposing. Additionally, since the disassembly of batteries is safety–critical,
this information also enables process automation.
The concept of a DT at battery EoL aligns with that of a Digital Battery Passport
(DBP). According to the European Commission legislation, all industrial batteries
and EVBs should implement a DBP by 2026 [43]. A DBP, as defined by the Euro-
pean Commission, is an electronic record (e.g., QR code) of an individual battery
containing crucial information about the sustainability characteristics of batteries,
such as capacity, performance, durability, and chemical composition. It aims to
support the scaling of circular economy strategies such as disassembly, repurposing,
remanufacturing, and recycling.

2.4.3 Implementation Framework

This use case falls under the category of DTI/DTA of a product, specifically the
battery system. When data from a population of battery systems is collected and
aggregated, it is referred to as a DTA. DTI/DTAs provide a method to aggregate the
entire lifecycle data of battery systems.
As illustrated in Fig. 7, data relevant to EoL decision-support is collected from all
other lifecycle phases. Design information from the DTP provides insights into the
original specifications and performance characteristics of the battery. Operational
data from the battery DTI/DTA is used to estimate SOH and predict RUL of the
battery in a second-life application. Promising methods exist for classifying end-
of-life batteries using physics-based models or data-driven models and algorithms.
Manufacturing data from the DTI offers details about the assembly format, materials,
and quality control processes used in producing the battery. The design-support tool
can use such information to identify inherent limitations or weaknesses in the battery
design that may impact its performance or safety during a second life.
The DTI/DTA at EoL facilitates the decision-making process to identify the most
suitable second-life applications. Similarly, the data can also be used to determine
other suitable pathways such as remanufacturing, repurposing, or recycling.
We propose the integration of DTIs as a solution for the DBP, providing a readily
available implementation. The technical design and operation of a DBP involve struc-
tured product-related data with a predefined scope, agreed data ownership, and access
rights conveyed through a unique identifier. By linking DTIs with DBP, stakeholders
Digital Twin in the Battery Industry 383

Fig. 7 DT at EoL

can access a wealth of information regarding the origin, usage, and disposal consid-
erations of batteries, thereby enhancing decision-making processes and promoting
sustainable practices.
DTIs contribute to a closed-loop lifecycle for batteries, supporting a circular
economy and responsible resource management. Their application promises signif-
icant environmental and economic benefits, transforming battery lifecycle manage-
ment toward sustainability and efficiency.

2.4.4 Challenges

. Similar to the usage phase, implementing DT at EoL also requires accurate

modeling of battery degradation, real-time data integration, and ensuring model
compatibility with diverse battery chemistries and applications. However, such
sophisticated modeling techniques are not commercially available.
. Managing uncertainties in predicting battery performance and health is another
obstacle, as uncertainties in degradation mechanisms, environmental factors, and
usage patterns can impact the accuracy of the predictions.
. Despite the potential benefits of integrating DTIs with DBP, the criteria for
EoL decisions remain heterogeneous and context-dependent. Standardizing these
criteria is essential for ensuring consistency and reliability in determining the EoL
status of batteries. This standardization process requires collaboration among
industry stakeholders, regulatory bodies, and academic researchers to estab-
lish universally accepted guidelines for assessing the suitability of batteries for
repurposing, recycling, or disposal.
To overcome these challenges, standardizing reference DTIs linked with DBP can
play a critical role. This approach ensures consistency and interoperability across
different stages of a battery’s life cycle. These reference models provide a common
framework for data collection, sharing, and analysis, enhancing transparency and effi-
ciency in tracking battery information. Standardization in the DBP can help overcome
384 S. Singh et al.

issues related to data silos, lack of interoperability, reliability concerns, and clarity
regarding responsibilities at various EoL stages. Ultimately, standardized reference
models play a crucial role in streamlining.

3 Case Study: Implementing Digital Twin of a Battery

System During the Operation Phase

3.1 Asset Administration Shell as Cornerstone

for Implementation of Digital Twin

In order to establish a general framework for implementing a DT covering the entire

lifecycle of a battery, as in our case, a standards-based approach is essential. This
is necessary to establish syntactic and semantic interoperability between different
actors or systems involved in the battery lifecycle. Such a standardized approach can
be found in the Asset Administration Shell (AAS), which we will closely examine
in this section.
Numerous terms and definitions introduced as part of Industry 4.0, such as Asset
Administration Shell (AAS) or DT, lacked a uniform understanding, even among
experts. Therefore, the focus of Wagner et al. [44] is not on defining additional
terms, but on explaining and substantiating existing terms with the aim of resolving
apparent contradictions. This takes the form of a classification of the AAS and the
DT in the lifecycle of a factory.
In their work, Tantik et al. [45] propose an approach to harmonize the spec-
ifications of the World Wide Web Consortium (W3C) with the guidelines of the
Plattform Industrie 4.0 (I4.0), thereby establishing a uniform structure for industrial
Cyber-Physical Systems (CPS). Using the recommended AAS for I4.0 components,
they identify and allocate required functionality to different segments.
The Plug-and-Produce concept necessitates the exchange of configuration data
when connecting a new module to a system. Lang et al. [46] suggest a novel concept
for secure Plug-and-Produce functionality, leveraging the combination of the AAS
and Blockchain technology.
In their work, Platenius-Mohr et al. [47] outline requirements and a solution for
enabling interoperable DTs by flexibly transforming their information models. They
illustrate the application of this solution in an industrial context by transforming IIoT
system-based DTs into the AAS format.
The AAS provides a standardized electronic representation of industrial assets
that act as DTs and enable interoperability between automated industrial systems
and CPS. Inigo et al. [48] present a case study on the application of AAS in an
industrial context, where the integration of a machine tool ecosystem with a robotic
arm was implemented to validate the AAS in a manufacturing scenario.
The AAS can also integrate existing standards from different domains or product
lifecycle phases. For example, Lüder et al. [49] discuss the capture and representation
Digital Twin in the Battery Industry 385

of engineering data as part of AAS. They present the AAS completion method based
on the AutomationML standard to facilitate the capture of the AAS engineering
dataset and its export to an AAS serialization.
The AAS can be considered from two perspectives: the virtual representation as
an asset description and the technical functionality as a smart manufacturing service.
In order to enable the effective provision of information for DT applications, Park
et al. [50] propose Virtual REpresentation for a DIgital twin application (VREDI),
which is an asset description for the operational procedures of a work-center-level
DT application.
These descriptions are aligned with the intended use of a DT, reflecting the
perspectives of involved professionals such as process designers, control special-
ists, and managers. Cimino et al. [51] introduce a paradigm, the Digital Multiverse,
to encompass major DT interpretations, not only in terms of data integration but also
by establishing and enforcing consistency rules involving both data and models.
As defined in the Reference Architectural Model for Industry 4.0 (RAMI 4.0), an
AAS represents a practical embodiment of the DT concept. Its realization involves
the integration of operational technologies and information and communication tech-
nologies. Ye et al. [52] aim to present the current status of AAS development, design
an intuitive method for implementing AASs, and develop an AAS-enabled digital
solution for cyber-physical applications in the manufacturing sector.
The survey by Jacoby et al. [53] of open-source implementations of the AAS
involved a structured search on GitHub and Google Scholar, which led to the selection
of four implementations for in-depth analysis. This paper represents a first attempt to
provide a comprehensive comparison of AAS implementations: ASX Server, Eclipse
BaSyx, FA3ST Service and NOVAAS.
In conclusion, it can be summarized that the AAS is now gaining traction after
many years of development and testing. Additionally, the AAS is open to integrating
existing standards and has the flexibility to cover all life cycle phases of an asset.
It can also be extended to design DTs for various application areas and domains.
Furthermore, a growing open-source community is also contributing to its widespread
adoption.

3.2 Digital Twin Concept Implementation with Asset

Administration Shell

As previously mentioned, the consideration of DT in the battery lifecycle is based on

use cases along this lifecycle, which are described using two DT concepts: the DTSM
and the DT types. How does the proposed DT implementation approach based on
the AAS relate to these two DT description concepts?
An AAS represents a DT implementation of an asset. An asset can be represented
by one or more AAS in virtual space, with each AAS specifically representing one
asset. An asset can encompass anything that holds value for a company, whether
386 S. Singh et al.

tangible or intangible. References to other AAS can also be managed within an AAS
in order to describe certain relationships or dependencies. Due to this nature of the
AAS, the DT types – DTP, DTI, DTA – can be implemented with the help of the
AAS.
The two dimensions of the DTSM, Lifecycle and PPR Model, can also be imple-
mented based on these characteristics of the AAS. Part of the third dimension of the
DTMS is already represented by the relationship between the AAS and the asset.
For the remaining cyber part (Descriptive and Executive), a deeper consideration of
the AAS metamodel is necessary. The AAS describes the aspects of a specific DT
of an asset using sub-models. These sub-models represent the semantic description
for such a DT aspect and consist of a collection of properties that describe certain
characteristics of the asset, such as the length, height, and width of a product or the
current temperature. These properties form the descriptive part. The executive part
consists of operations that can be defined as part of a sub-model. Operations take
properties as input parameters and return properties as results. In this way, executable
functions and services can be described and called.
The actual execution of an operation is implementation specific; for instance, a
simple arithmetic operation or a simulation model can be executed.
For this reason, the AAS is also ideal for implementing a DT for the battery life-
cycle, as discussed in more detail in the following sections through implementation
examples.

3.3 Case Study: DT Implementation for State Estimation

of LiB Cells

Building upon the understanding of the AAS, we will now examine how it serves
as an ideal framework for implementing a battery DT during the operation phase to
estimate the state of the battery. In this case study, a battery DT software application
is implemented, which can be installed and accessed by various stakeholders, such
as battery specialists seeking insights into the cell behavior in actual applications
during usage. Models and battery data are associated with the assets, thereby serving
as the backend of the software application, which is then used for practical use cases.
There is no one-size-fits-all approach for implementing this software application
because its usage heavily depends on the IT infrastructure of the application land-
scape. Therefore, a tailored implementation strategy is necessary to accommodate
the unique characteristics and requirements of each application environment.
Drawing inspiration from the foundational principles outlined in [54–56], a
DTI for state estimation of LiBs is implemented, and the architecture of the
implementation is presented in Fig. 8.
The DTI components are as follows:
1. Asset: This refers to the real-world entity or system that the DT aims to replicate.
For this implementation, the asset is a battery pack. Please note that an asset of
Digital Twin in the Battery Industry 387

Fig. 8 DTI architecture for state estimation of LiBs using AAS

a cell or module can be referenced to the AAS of the battery pack. This also
implies that the DTI of a battery pack is a DTA of the battery modules.
The AAS, implemented through the AASX package explorer, is shown in Fig. 9. It
consists of a number of sub-models in which all the information and functionalities of
the battery – including its features, characteristics, properties, statuses, parameters,
measurement data, and capabilities are stored. The granularity of operation data fed
into the DTI depends on the interface responsible for real-time data transfer from the
BMS.
By creating an AAS for the battery pack, we have established a standardized
framework for accessing, managing, and exchanging data related to the battery’s
operation, maintenance, and performance. This AAS enables seamless integration
of operation data with the models, facilitating holistic monitoring, analysis, and
optimization of the battery pack’s usage and health. Through the AAS, stakeholders
can gain valuable insights into the battery’s state, make informed decisions, and drive
continuous improvement in its performance and reliability.
2. Data Collection and Device Control Entity: The AAS refers to this entity, which
comprises two sub-entities. The first sub-entity is responsible for collecting data
from different sources. Battery-related data such as design data, experimental
data, operation data, data derived from simulations, and expert knowledge about
battery behavior form the basis data for battery DT. These data are used for model
development, verification, evolution, and execution. It is important to note that
388 S. Singh et al.

Fig. 9 AAS of a battery

the methods/interfaces used for acquiring and storing these diverse data sets may
vary. The second sub-entity is the control unit, like the BMS, which sends control
programs to the battery system when adjustments are needed.
3. Core entity: The core entity acts as the “brain” of the DTI, serving as a cognitive
center. The functions performed by the core entity include data management,
data analytics, model development and management, model verification, result
generation, and model evolution. Additionally, the interfaces that connect the
software components are also part of the core entity.
The data management module processes and manages all types of data acquired
in the data collection entity. In the model management module, all models can
be stored, used, managed, and updated. The DT battery models include physics-
based electrochemical models, ageing models, data-driven model lifecycle prediction
models, etc. The algorithms responsible for generating results (estimations, predic-
tions, or forecasts, and battery behavior evaluation) from the models or the data are
also integrated in this module. As the battery is used the DT accumulates data from
the usage phase such as charging/discharging profiles, working environments sensor
acquisitions, and maintenance data. By combining historical and real-time data, the
DTI parameters can be updated and evolved using methods such as neural networks,
Bayesian regression, maximum likelihood parameter estimation, and other param-
eter estimation algorithms. This model evolution module works continuously to
enhance the accuracy of the DTI model.
4. User Entity: The DT software application is a web service accessed by the user.
This application should include some type of visualization module for users to
access the results of the DTI. For this implementation, we utilized the AAS Web
GUI [57]. AAS Web GUI is a user interface provided by Basyx for interacting
with the AAS Server and AAS Registry through a web browser. For further details
Digital Twin in the Battery Industry 389

Fig. 10 AAS GUI of the battery DTI

on asset integration through Basyx, readers are referred to [58]. Figure 10 shows
a sample view of the AAS GUI for the battery pack.
The end user of the DTI is not necessarily the end user of the battery system as
well. The DT developer is responsible for accessing the core entity and its connection
with the other entities.
Implementing a DT for state estimation of LiBs cells holds significant promise
in enhancing our understanding of its degradation mechanisms and in optimizing its
usage in applications. The AAS framework enables an interface for bi-directional
data exchange between the physical and the virtual space. However, a limitation
remains that the accuracy of the DT heavily relies on the quality and granularity
of data fed into the battery models. This challenge can be mitigated by employing
reduced-order and adaptive models, which offer possible ways to improve the model’s
robustness. Despite these challenges, a DT implementation represents a significant
step forward in advancing battery management practices. By continually refining
methodologies and overcoming limitations, we can unlock the full potential of DTs
in revolutionizing the way we monitor, analyze, and optimize LiBs for improved
performance and reliability in real-world settings.

4 Conclusion

Across the sections of this chapter, we discussed the use cases in different lifecycle
phases of a battery, illustrating the implementation framework and the corresponding
challenges encountered during implementation. We shared insights into the prac-
tical considerations and strategies necessary for the implementation. Moreover, we
described an example of our own implementation, offering a tangible demonstration
390 S. Singh et al.

of how the theoretical concepts discussed throughout the chapter can in applied in a
in real-world battery applications.
In conclusion, this chapter highlights the transformative potential of DTs in revo-
lutionizing decision-making and operations throughout the battery lifecycle. From
real-time monitoring to predictive maintenance and continuous data-driven insights,
digital twins emerge as powerful catalysts for efficiency, innovation, and sustain-
ability in the dynamic landscape of battery technology. The chapter underscores the
profound impact of the DT concept on advancing capabilities and environmental
considerations associated with batteries across various applications, setting the stage
for a paradigm shift in how we perceive and interact with battery systems.

References

1. VDI/VDE Innovation + Technik GmbH: Projektträger und Dienstleister für Innovationen

(2024) Battery cell manufacturing ecosystem in Europe. https://vdivde-it.de/en/publication/
battery-cell-manufacturing-ecosystem-europe. Accessed 25 Mar 2024
2. Wirtschaft und Klimaschutz, BMWK—Bundesministerium für (2024) Batterien für die Mobil-
ität von morgen. https://www.bmwk.de/Redaktion/DE/Artikel/Industrie/batteriezellfertigung.
html. Accessed 25 Mar 2024
3. Naseri F, Gil S, Barbu C et al (2023) Digital Twin of electric vehicle battery systems: compre-
hensive review of the use cases, requirements, and platforms. Renew Sustain Energy Rev
179:113280. https://doi.org/10.1016/j.rser.2023.113280
4. Deng J, Bae C, Denlinger A et al (2020) Electric vehicles batteries: requirements and challenges.
Joule 4:511–515. https://doi.org/10.1016/j.joule.2020.01.013
5. Masias A, Marcicki J, Paxton WA (2021) Opportunities and challenges of lithium ion batteries
in automotive applications. ACS Energy Lett 6:621–630. https://doi.org/10.1021/acsenergylett.
0c02584
6. Link S, Neef C, Wicke T (2023) Trends in automotive battery cell design: a statistical analysis
of empirical data. Batteries 9:261. https://doi.org/10.3390/batteries9050261
7. Stephan A, Hettesheimer T, Neef C, et al (2023) Alternative battery technologies roadmap
2030+. Fraunhofer-Gesellschaft
8. Link S, Neef C, Wicke T, et al (2022) Development perspectives for lithium-ion battery cell
formats. Fraunhofer ISI
9. Baars J, Domenech T, Bleischwitz R et al (2021) Circular economy strategies for electric
vehicle batteries reduce reliance on raw materials. Nat Sustain 4:71–79. https://doi.org/10.
1038/s41893-020-00607-0
10. Michaelis S, Schütrumpf J, Kampker A, et al (2023) Roadmap battery production equipment
2030. Update 2023. VDMA Verlag
11. Kampker A, Heimes HH, Offermanns C et al (2023) Prediction of battery return volumes for
3R: remanufacturing, reuse, and recycling. Energies 16:6873. https://doi.org/10.3390/en1619
6873
12. Thielmann A, Wietschel M, Funke S, et al (2020) Batterien für Elektroautos: Faktencheck und
Handlungsbedarf. Sind Batterien für Elektroautos der Schlüssel für eine nachhaltige Mobilität
der Zukunft?[Batteries for electric cars: Fact check and need for action. Are batteries for electric
cars the key to sustainable mobility in the future?]. Perspectives—Policy Briefs
13. Doose S, Mayer JK, Michalowski P et al (2021) Challenges in ecofriendly battery recycling
and closed material cycles: a perspective on future lithium battery generations. Metals 11:291.
https://doi.org/10.3390/met11020291