PreprintPDF Available

Digital Twins: State of the Art Theory and Practice, Challenges, and Open Research Questions

Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

Digital Twin was introduced over a decade ago, as an innovative all-encompassing tool, with perceived benefits including real-time monitoring, simulation and forecasting. However, the theoretical framework and practical implementations of digital twins (DT) are still far from this vision. Although successful implementations exist, sufficient implementation details are not publicly available, therefore it is difficult to assess their effectiveness, draw comparisons and jointly advance the DT methodology. This work explores the various DT features and current approaches, the shortcomings and reasons behind the delay in the implementation and adoption of digital twin. Advancements in machine learning, internet of things and big data have contributed hugely to the improvements in DT with regards to its real-time monitoring and forecasting properties. Despite this progress and individual company-based efforts, certain research gaps exist in the field, which have caused delay in the widespread adoption of this concept. We reviewed relevant works and identified that the major reasons for this delay are the lack of a universal reference framework, domain dependence, security concerns of shared data, reliance of digital twin on other technologies, and lack of quantitative metrics. We define the necessary components of a digital twin required for a universal reference framework, which also validate its uniqueness as a concept compared to similar concepts like simulation, autonomous systems, etc. This work further assesses the digital twin applications in different domains and the current state of machine learning and big data in it. It thus answers and identifies novel research questions, both of which will help to better understand and advance the theory and practice of digital twins.
Content may be subject to copyright.
Digital Twins: State of the Art Theory and Practice, Challenges, and Open
Research Questions
Angira Sharma, Edward Kosasih§, Jie Zhang§, Alexandra Brintrup§, Anisoara Calinescu
Department of Computer Science, University of Oxford
§Institute for Manufacturing, University of Cambridge
Digital Twin was introduced over a decade ago, as an innovative all-encompassing tool, with perceived benefits
including real-time monitoring, simulation and forecasting. However, the theoretical framework and practical imple-
mentations of digital twins (DT) are still far from this vision. Although successful implementations exist, sucient
implementation details are not publicly available, therefore it is dicult to assess their eectiveness, draw compar-
isons and jointly advance the DT methodology. This work explores the various DT features and current approaches,
the shortcomings and reasons behind the delay in the implementation and adoption of digital twin. Advancements in
machine learning, internet of things and big data have contributed hugely to the improvements in DT with regards to
its real-time monitoring and forecasting properties. Despite this progress and individual company-based eorts, cer-
tain research gaps exist in the field, which have caused delay in the widespread adoption of this concept. We reviewed
relevant works and identified that the major reasons for this delay are the lack of a universal reference framework,
domain dependence, security concerns of shared data, reliance of digital twin on other technologies, and lack of quan-
titative metrics. We define the necessary components of a digital twin required for a universal reference framework,
which also validate its uniqueness as a concept compared to similar concepts like simulation, autonomous systems,
etc. This work further assesses the digital twin applications in dierent domains and the current state of machine
learning and big data in it. It thus answers and identifies novel research questions, both of which will help to better
understand and advance the theory and practice of digital twins.
1. Introduction
How can one reduce the cost of producing a prototype and performing tests on it? How can one perform the
extreme tests on a prototype which cannot be performed in a laboratory? How can a prototype imbibe all the infor-
mation and outcome of these tests, to provide an accurate prediction to the future experiments? How can one monitor
a physical asset in real-time1and be alerted before anything goes wrong? How can we humans have access to this
real-time information of all the components involved in a physical asset, perform meaningful real-time analysis on
this information, and make timely, robust and ecient decisions for future operations based on them? The answer –
Digital Twin (DT).
Digital Twins are virtual product models which encompass all the above qualities [65]. Grieves and Vickers [29]
define the Digital Twin (DT) as ”a set of virtual information constructs that fully describes a potential or actual physical
manufactured product from the micro atomic level to the macro geometrical level. At its optimum, any information
that could be obtained from inspecting a physical manufactured product can be obtained from its Digital Twin. Digital
Twins are of two types: Digital Twin Prototype (DTP) and Digital Twin Instance (DTI)”. DT aims to combine the
best of all worlds, namely, twinning, simulation, real-time monitoring and analytics. Digital Twin has been recognised
as the next breakthrough in digitisation, and also as the next wave in simulation [73,62]. It can save cost, time and
resources for prototyping, as one doesn’t need to develop the physical prototype(s), but can perform the same tests
1Real-time means at the time frequencies at which the state of a physical asset changes or is expected to change significantly. For example: for
an aircraft it could be in seconds, minutes or hours, whereas for a manufacturing unit it could be in hours or days.
arXiv:2011.02833v2 [cs.LG] 6 Nov 2020
on a virtual prototype without aecting the real operation [32,63]. Gartner, a research and advisory company, listed
Digital Twins as one of the ”Top 10 Strategic Technology Trends” for 2017, 2018, and 2019 [10,11,22]. Furthermore,
Market Research Future predicts that the Digital Twin market will reach 35 billion USD by 2025 [47].
Despite the stated advantages and potential of DT technology, certain research and implementation gaps exist,
which have hindered the adoption and advancement of this concept since its inception around 2003: the diverse
applicability of DT is a result of its reliance on advanced evolving technologies, mainly IoT, big data and machine
learning. The real-time monitoring and data collection capability of a DT is reliant on IoT devices in the environment
and enterprise information systems, whereas the analytics is reliant on big data and machine learning tools. Combining
these technologies and implementing them for one or more physical assets requires extensive domain knowledge, as
would be required while creating any physical asset’s physical prototype. Furthermore, the DT vision is continuously
evolving, as the technology, industry and customer needs evolve. Not only is the overall concept of DT not fully
established, there is no universally accepted definition of DT, and no established standards for implementing it. As
a technology which is used across various domains and is dependent on other evolving technologies, it eventually
becomes dependent on the current state of these technologies and has to be tailored for each domain as well. Having
no standards further impedes the widespread design, implementation and adoption of this technology.
In this work we attempt to study these gaps and propose solutions. As DT is a practical concept, this work
emphasises the need for a comprehensive theoretical specification – via a DT reference framework, leading to an
implementation and evaluation methodology. The contributions of this paper are:
1. Assess how the dependence on other technologies is impeding the advancement and adoption of the DT tech-
nology and study the involvement of machine learning and big data in it.
2. Study how domain (or sector or type of industry) aects the DT implementations.
3. Study the gap between the ideal DT concept and practical implementation.
4. Evaluate the current state of DT and discuss and answer its limitations and challenges.
5. Provide a comprehensive analysis of the past eorts at theoretically establishing, and practically implementing
digital twins.
6. Deduce the answer to the debate concerning the definition and implementation of DT.
Assumptions: Due to lack of documented implementation methodology of DT, it is infeasible to fully comprehend
how the various components such as IoT, data, simulation, real-time synchronisation and machine learning aect the
implementation of DT. Keeping in mind the importance of these components for the concept of DT, in this work we
assume good inter-connectivity and compatibility among all these components.
The rest of the paper is organised as follows: Section 2discusses and contrasts the existing reviews. Section 3
discusses the concept, components, properties, versions, evolution and assumptions of DT. Section 4discusses the
existing theoretical, practical and industrial models of DT across dierent domains. It also explores the existing use
and potential of machine learning and data in DT. Section 5reviews the industries where DT can be particularly
useful. Section 6discusses the challenges and limitations with the existing DT models. Section 7discusses the future
directions and poses some new questions for the field, and Section 8concludes the paper.
2. Previous Work
As DT is a general concept that can be tailored for a particular domain, there has been a decent amount of
literature reviews focusing on either its implementation in specific domains or in general. Some works focus on
specific domains, such as manufacturing [54], aerospace [60] and production science [39], others are general [20,73].
Most reviews assess the current definition of DT and provide new insights. For example, Kritzinger et al. [39]’s
categorical review focuses on production science and categorically lists the papers in the areas of Digital Twin, Digital
Shadow and Digital Model, whereas Negri et al. [54] explores all the papers in industrial manufacturing, mainly
dierentiating whether big data and data models exist in the literature. Consequently, the insights of these reviews are
domain-specific, rather than generic and transferable across domains. For example, [39] looks at similar concepts of
Digital Shadow and Model from a manufacturing perspective, whereas [54] explores data models in a specific class
of manufacturing systems. [20] provides a general view on DT. It compares across 87 applications and proposes
a classification scheme. However, although such reviews are extensive, they do not include any white papers or
industrial works. As DT is both a theoretical and applied concept, leaving out the practical implementations would
provide an incomplete and insuciently representative description, hence we include these in our review. Tao, Zhang,
Liu and Nee [73] are among very few authors who explore the specific practical implementations of DT.
The previous works build very little on practical implementations, and focus more on the categorisation of papers
and development of theoretical models. The reasoning behind this, as mentioned in the Introduction and further elabo-
rated in Section 4.2, is that, because DT is domain-dependent and rely on multiple technologies, the conceptualisation
of DT – taking the domain into regard and coming up with a suciently-detailed and specified, yet universal model
– is a dicult task. Nonetheless, these reviews pose the following questions and insights, which help to advance the
DT concept further. In this review we discuss these pressing issues in DT:
The need for a proper definition of DT which could cover the dierent application domains [46,73].
An investigation of concrete case-studies, and a reference model for domain-specific requirements [39,54].
An investigation on how DT research could contribute to the Internet of Things (IoT) and Information Systems
(IS) disciplines [20].
The analysis of the debate initiated in 2013 [40] and further explored in [54], whether DT covers the entire
product lifecycle [62,23] or just the product [66,1].
The scope and need of Big Data and Data Models in DT [54], and data-related issues [62].
Apart from the above, in this paper we pose some new research questions and insights, which aim to help exploring
and advancing the practical applications of DT.
2.1. How this review diers
Previous reviews identified critical DT insights and questions. This review answers them and poses some new
questions, to further advance the DT theory and practice. Also, this review presents the limitations, challenges, and
new properties of DT, by researching through the practical implementations of DT, which was previously under-
explored. Compared to the existing reviews, this review does not focus on the categorisation and statistics of the
literature based on models, domain, etc., but rather presents the conceptual view of the implementation of DT concept
by looking at the root causes behind the challenges of implementing DT. Also, we address the following novel aspects
and questions:
1. A universal reference framework for DT.
2. The gap between the actual DT concept and implementation (due to no universal definition or architecture).
3. The importance of real-time machine learning and data in DT.
4. What is impeding the spread of DT?
5. How domains aect the implementation of DT?
6. How successful the existing implementations of DT are?
In all, this review provides a dierent viewpoint and analysis to the existing literature and DT implementations.
3. Digital Twin
Digital Twins were first introduced in early 2000’s by Michael Grieves in a course presentation for product life-
cycle management [28]. In 2011, implementing DT was considered a complex procedure, which required many
developments in dierent technologies [77]. Despite being coined in 2003, the first description to use of Digital Twin
was years later by NASA in Technology Roadmaps [67], where a twin was used to mirror conditions in space and
to perform tests for flight preparation (another example of such hardware twin was the Airbus Iron Bird). Dawned
with the aerospace industry, soon around 2012 the manufacturing industry started to use DT. So what took the concept
nearly a decade to be implemented?
With the advancements in technologies like cloud computing, IoT and big data, many domains have seen major
developments, such as Industry 4.0 [44], Physical Internet [4,70], Cyber-manufacturing [37], Made in China 2025
[41], Cloud Manufacturing [84] etc. Industry 4.0 has seen a revolution mainly because of digital advancements,
IoT and Big Data [54,36]. It was because of the Industry 4.0, the storage of all data in digital format, and sensors
being inbuilt into the industrial spaces, that the implementation of digital twin was possible, rejuvenating the concept.
Moreover, with the emerging extensive simulation capabilities, it became feasible to perform realistic tests in a virtual
environment. Owing to these technical advancements, the idea of implementing a functional DT was soon adopted by
companies like IBM, Siemens and GE, as a utility for themselves and for their clients.
3.1. Defining DT
On the journey to define the model and properties of DT, various terms have existed in multiple literature works,
such as- ’ultra-high fidelity’ [59] , ’cradle-to-grave’ [76], ’integrated’ [76] model , ’integral digital mock-up (IDMU)’
[60]. These terms are important and relevant to the DT concept, however, having multiple definitions and terms has
delayed reaching a consensus on a single representative, unifying definition.
In the simplest words, a digital twin is a ’digital’ ’twin’ of an existing physical entity. What makes a DT all of the
above are its properties, which we discuss ahead.
Though the literal meaning of digital twin seems simple at first, the definition of DT has been a subject of debate.
For some members of the community (Schroeder et al. [66], Abramovici et al. [1]) it is correct to consider the Digital
Twin as the final product, whereas for others (Rosen et al. [62], Gabor et al. [23]) it is the entire product lifecycle.
Note that, for consistency and generality, we have called the conventionally termed ’product’ an ’asset’ in rest of
the paper.
3.2. Components of DT
DT was first introduced by Grieves [28] with three components: the digital (virtual part), the real physical product,
and the connection between them. However, other authors, such as Tao et al. [72] have extended this concept to have
five components, by including data and service as a part of DT. Tao et al. [73] also identify VV&A (verification,
validation and accreditation) as DT components. With data models coming into the picture, Miller et al. [52] extend
the definition of DT to be an integration of multiple models of a model-based enterprise (by creating associations
between dierent models and relations between data stored in dierent parts, a digital twin can be formed).
As conceptually sound the above definitions are, reaching consensus on a DT definition requires specifying the
fundamental requirements for a DT. With the advancements in technologies on which DT depends (such as machine
learning, big data and cybersecurity) these requirements have changed over time. Moreover, the domain-dependence
of DT call for a defining the components which can be generalised across domains.
There have been components and properties which have existed in some works and have been missing in others.
We collate this information in the Table 1. The properties and components which we state are necessary, are a result
of their existence in the literature, and our understanding of the DT concept. We thus integrate the contributions of
previous works, which have only been concerned with some components of DT, to provide a holistic definition of DT.
Based on this analysis and our understanding, we define the elementary and imperative components of a DT.
3.2.1. Elementary Components
The elementary components are those without which a DT cannot exist:
1. Physical Asset (could be either a product or a product lifecyle)
2. Digital Asset (the virtual component)
3. 2-way synchronized relation between the physical and digital asset (for information flow)
3.2.2. Imperative Components
The imperative components add to the properties of DT, to make it the all-encompassing tool of simulation, real-
time monitoring and analytics. Without any of these, the uniqueness of DT ceases to exist:
1. IoT devices – to collect real-time information from various sub-components of the physical asset.
Requires: High-fidelity connection between all IoT devices, for accurate and timely flow of information.
2. Integrated Time continuous Data – gathered from dierent IoT components for machine learning and analytics is
essential in order to monitor the system, guarantee correct behaviour and provide input to the machine learning
Requires: Big data analysis and storage tools for extracting useful information from data.
3. Machine learning– for predictions and feedback as well as to propose and mitigation strategies, in exceptional
Requires: A joint optimisation feature for all the sub-components of the DT.
4. Security of the data and information flow among the various physical components of the physical asset.
5. DT Performance evaluation.
Requires: Evaluation metrics (e.g. accuracy, resilience, robustness, costs), and evaluation methods and
Table 1: Adjacency matrix showing the various components present and absent in literature [’’ indicates present and ’ ’ indicates absent, ’*’ means
indication but not explicitly]
Papers Components
of Info.
IoT Static
ML Domain
Testing Security
[67]• • • •*
[77] • • *• •*• •*
[26]• • • •*
[60] • •
[62]• •*• •
[28] • • • •
[66] • •
[65] • •
[54] • •
[39] • •
[72] • • • •
[18]• • *
[42] • • *• •*• •*
[73] • •* • •
[20]• • * * *
[53]• • • •*
[14] • • * • •*
[58] • • * • • **
[30] • • *• •**
[46] • • *• •
Table 2: The dierent components of DT and their key roles
Component Role
Physical Asset what the digital twin is a twin of
Digital Asset the digital twin
Continuous Bijective Relation for real-time synchronisation and twinning
IoT for data collection and information sharing
Time continuous data for synchronisation and input to machine learning
Machine learning for analytics of the asset
Security to prevent data leaks and information compromises
Evaluation metrics /Testing to evaluate the performance of DT
Table 2summarises how each component contributes uniquely to the functions of DT. Removing any component
voids the DT of the functionality and its uniqueness, as discussed ahead.
3.3. Properties of a DT
As simple as the definition sounds, the properties of the DT are what makes it more than just a ’digital’ ’twin’.
3.3.1. Necessary Properties
These are the properties inherent to any digital twin:
1. Real-time connection with the physical entity ([60] defines a ’biunivocal’ relation between DT and the physical
2. Self-evolution: A very important property which has not been explored much (introduced in [72]). With this
DT can learn and adapt in real-time, by providing feedback to both physical asset (via the human asset) and DT.
This can be easily harnessed now due to the uprise of machine learning tools: to remodel and redesign itself
(such as reinforcement learning). The frequency of this synchronisation depends on the update scenarios, such
as event-based (supply chain), periodic intervals (aircraft), condition based (logistics), etc.
3. Continuous machine learning analysis (dependent on the frequency of the synchronisation), not just one-time
output forecasting.
4. Availability of time-series (or time continuous) data for monitoring [72], and as input to machine learning
5. Domain dependence (or Domain specific services): According to the domain, a DT may provide or prioritise
services specific to the industry. These are the same ’domain specific’ services which exist in the physical asset
(for example the optimisation problems for an aircraft and a manufacturing unit will prioritise or add more
weight to dierent parameters).
3.3.2. Dynamic Properties
Based on these dynamic properties, a hierarchy of digital twin can be created:
Autonomy: A DT (or for that matter any information provider [13]) could either make changes to the physical
asset itself, or a human in control could make changes to the DT. This applies dierently to dierent hierarchies
of components present in the twin, such as to some parts of the machine learning system, or some part of the
decision making system. Hence, the property of a DT to be autonomous, not autonomous, or partly autonomous
is case-dependent.This classification also includes the self-evolution mechanism of DT (what changes must it
make to itself, and what changes must be approved by a human).
Synchronisation: Synchronisation of data could either continuously or at certain time intervals. These depend
on a number of factors such as technology, resources available, need for the data and type of machine learning
algorithm being used. A DT could have sub-components which could be partly continuously synchronised and
partly event-based synchronised.
This synchronisation can result in dierent hierarchies based on the following:
-How often the data is collected?
-How often the data is stored?
-How often the DT is updated? (note that this is dierent to the property of autonomy as that deals with
’who’ rather than ’when’)
Based on the information present in current literature and our analysis, we present a reference framework in Figure 1.
Answering the debate concerning whether a DT should be a product or product lifecycle, after analysing the literature
our take on it is that, a DT can be both, for example, DT of a car or DT of the car’s production lifecycle. This stems
from the components of DT defined in Section 3.2; as long as all the components are present, any physical asset could
have its Digital Twin.
Figure 1: DT reference framework with components and information flows:
Flow of Information - IoT, Data and Machine Learning Data will be gathered from various IoT devices and sent to the DT. DT will use this data
for machine learning and analytics. The prediction from DT will be provided to the human, to take further steps. Details of these further steps
could be fed back into DT, to test how their corresponding actions will aect the DT.
3.4. How is DT dierent from existing technologies
The diverse applications of DT such as simulation, real-time monitoring, testing, analytics, prototyping, end-to-
end visibility [35], can be broadly classified as sub-systems of DT (for example, a DT can be used for testing during
prototyping, for real-time monitoring and evaluation, or for both). It is the presence of all the components discussed
in the previous section that makes a DT dierent from these, as described in Table 3.
Table 3: How DT diers from existing technologies
Technology How it diers from DT
Simulation No real-time twinning
Machine Learning No twinning
Digital Prototype No IoT components necessarily
Optimisation No simulation and real-time tests
Autonomous Systems No self-evolution necessarily
Agent-based modelling No real-time twinning
3.5. A brief overview of other similar concepts that preceded DT
The concept of digitising and twinning are not new. Many similar concepts have preceded DT, however, for the
reasons described briefly below they dier.
Digital Shadow, Digital Model - [39] A Digital Model has only manual exchange of data and it does not
showcase the real time state of the model. Digital Shadow is a saved data copy of the physical state, with one-
way data flow from physical object to the digital object. The digital twin on the other hand, has fully integrated
data flow where the digital twin properly reflects the actual state of the physical object.
Semantic Virtual Factory Data Model (VFDM) are virtual representations of factory entities [75]. These were
used in manufacturing and industrial spaces [74]. DT diers from VFDMs because of the property of real-time
synchronization. On one hand, where VF is a data model only, on the other, DT is real-time and synchronised.
Product Avatar is a distributed and decentralised approach for product information management with no con-
cept of feedbacking and may represents information of only parts of the product [60]
Digital Product Memory: [52] see DT as an extension of semantic /digital product memory where a digital
product memory senses and captures information related only to a specific physical part and thus can be viewed
as an instantiation of DT.
Intelligent Product [48,83,50]: A DT can be seen as an extension of Intelligent Product which uses the new
technologies like IoT, Big Data and Machine Learning which were lacking in the concept of Intelligent Product.
Holons [49,79,78]: As an initial computer-integrated manufacturing tool, holons formed the basis for all the
technologies described above.
Clearly DT has many novel and exclusive collection of applications to provide. Having established its background
and components, we now process to its models in the next sections.
4. Existing Models
4.1. Theoretical models
In this section we explore the various attempts to define the architecture of DT. Most of these are for particular
domains/sectors, while few are generalised.
The most popular sector for implementing a DT is Product Lifecyle Management (PLM). This is chiefly because
DT provides a holistic view to the widespread components of PLM which is useful to solve the existing problems in
this domain [38]; although challenges persist, such as Tao et al. [72] discuss how data in PLM is isolated, fragmented
and stagnant which poses issues. The work also presents a theoretical framework of applying DT to PLM by proposing
three design steps: conceptual design, detailed design and virtual verification. The work gives an example of DT for
bicycle as a case study, but does not mention the implementation methodology. Such as to perform tests on the
bike with parameters involving brakes, speed, customer weight, etc., proper physics simulators are needed and these
simulation softwares or frameworks have to be defined.
Liu et al. [42] propose using Unity3d open source engine for implementing DT. They create a reference model to
handle and synchronize complex Automated Flow-Shop Manufacturing System (AFMS) systems using DT and deals
with decoupling multi-objective optimization problems. As mentioned in Section 3, if DT is an entire product lifecy-
cle, then analytics and using machine learning in DT would mean defining a universal joint optimization problem for
the lifecyle. Handling multitude of parameters, and solving multi-objective optimisation problems is an existing chal-
lenge for complex systems. However, Liu et al. [42] apply the sheet processing DT prototype successfully to Chengdu,
China, and use heuristic measures such as production packages, unified system cost, unified system performance to
evaluate the performance of DT (resulting in a domain-dependent DT).
This domain dependence property is also discussed in [20], where a generalised architecture of DT is presented
by defining six dimensions for defining the concept of DT – industrial sector, purpose, physical reference object,
completeness, creation time, and connection (some of which were inspired from previous works). The other important
self-evolving property of DT is explored in [72], where the authors discuss many cases of using DT in product design
and manufacturing.
Continuing the debate on whether DT should be the product or the product lifecycle, Schleich et al. [65] propose a
DT reference model for production systems, for geometrical product specification (GPS), where they pitch the idea to
represent the DT as an abstract form of the physical product, mentioning clearly that DT is the entire product lifecycle
and not just the product.
As Digital Twins can be considered as artificially intelligent systems, they are closely linked to autonomous
systems. This link is explored in [62], where the authors discuss the importance of DT in simulating the autonomous
system’s decisions. As autonomous systems need real-time information, the paper argues that digital twin will be very
useful and ecient for the autonomous systems to function properly, as DT can collect all data and prior knowledge
needed. They discuss the driving aspects for the future of manufacturing, which are: Digital Twin, modularity,
connectivity and autonomy. These aspects are also important for the concept of DT – to see the individual components
of DT and then integrate them into a single model. The authors also support the claim that DT is the next wave in
simulation, as it encompasses real time operation and data. Despite of a well-explored evaluation, the examples
considered in the [62] paper for implementing a real DT are limited to using DT as a memory buer to draw data, and
do not explore the machine learning and analytics potential.
Despite being conceptually sound and well-explored, the current DT literature does not fully address the gap
between theory and practice, due to the challenges which we discuss ahead in Section 6.
4.2. Practical models and their domain-dependence
The applications of DT spans across various domains from manufacturing, aerospace to cyber-physical systems,
prognostics, health and management, etc. The description of these practical models in literature is scanty. Moreover,
eventually, as hard and complex implementing a DT for large systems becomes, these descriptions become vague.
Not to forget, this complete methodology is domain dependent as well. We discuss these practical attempts and the
aect of domain on these attempts ahead.
Owing to the number of parameters and intricacies of the operations involved the most complex but most bene-
ficial implementation of a DT is in the aerospace domain. [26] discusses application of DT for NASA and U.S. Air
Force vehicles. The main advantage of using a DT for aerospace is that a DT can replicate the extreme conditions
(thermal, mechanical and acoustic loadings) which cannot be physically performed in a laboratory because laboratory
tests cannot go below or above a certain limit. Moreover, conventional approaches in this domain do not consider the
specific material involved and type of the component during testing, which is highly desirable. On the other hand, a
DT can be tailored according to tail number to take into account the specification of the materials and its types. Ad-
ditionally, the current analytics approaches are mostly based on statistical distributions, material properties, heuristic
design properties, physical testing and assumed similarities between testing and real operation. But, as the authors
say, these techniques are unable to represent the future extreme requirements, whereas the DT can use maintenance
history and other historical fleet data to continuously forecast health and probability of mission success. A DT can
also test the future extreme requirements in real time, as and when they appear. This real-time support is needed as
external support is not always possible. Also, the machine learning capabilities of a DT can make predictions and
recommendations for in-flight changes to a mission. A DT also provides self-mitigating mechanisms or recommen-
dations. It is because of the plethora of advantages of using a DT, mainly as an alternative form of testing which
can’t be performed in the lab, that DT in aerospace is highly desirable, nonetheless, hard to implement because of the
multitude of parameters involved.
ıos et al. [60] presents a review of using DT in aerospace where the authors describe that, because a commercial
aircraft may have more than half a million dierent component references, it is challenging to create a bijective
relation between a particular physical aircraft and its unique digital twin. The complexity in implementation of DT
also arises due to the interoperability issues among the dierent softwares used in production, such as PLM, Enterprise
Resource Planning (ERP) , Manufacturing execution system (MES), Computer-aided technologies (CAx), etc. The
work suggests to make use of each product’s individuality such as Manufacturers Serial Number (MSN), the EPC
(electronic product code), Tail Number (TN), aircraft registration, VIN (vehicle identification number) to create an
eective DT, which is similarly dicult because of the number of parameters involved.
On the other hand, Tao et al. [73] discuss how prognostics and health management (PHM) have the biggest
advantage of using DT as the concept considers behaviour and rules modelling, ultrafidelity and merges physical,
virtual, historical and real-time data to provide trends, optimization and maintenance strategy. The work also discusses
the challenges pertaining to the cyber-physical fusion which occurs while implementing DT; these challenges are
security, robustness, applicability, data acquisition, mining and collaborative control.
The advantages and challenges of using a DT are thus dierent for dierent domains. Therefore, a DT has to
be tailored according to these dierent sectors. These dierences are explored in [54]. This work discusses the
history of DT and highlights briefly the expectations from DT in dierent areas such as CPS (focus is on avoiding
failures, support health analysis of systems and deformation of materials in the physical twin, and study the long
term-behaviour in dierent environments), aerospace (maintenance and intervention needs of the aircraft with the
use of Finite Element Methods (FEM), Computational Fluid Dynamics (CFD), Monte Carlo and Computer-Aided
Engineering (CAE) applications-based simulations), manufacturing (to simulate complex and numerous parameters
of the system) and robotics (to optimise control algorithms during development phase). These highlight how a DT is
expected to behave dierently as best suited for the dierent domain.
To the best of our knowledge, the use of ML techniques in DT is next to none. Only [53] lays down a proof of
concept for implementing DT with machine learning, in the petrochemical industry. The paper provides some time
series data preprocessing solutions for unifying frequency of time series data, resolving time lag issues between time
series data, reducing data dimensions, regenerating new time series data, etc. They have provided few implementation
details: the authors use many production service systems for data gathering and training the digital twin (such as super-
visory control and data acquisition (SCADA), programmable logic controller system (PLC), manufacturing execution
system (MES), laboratory information management system (LIMS), distributed control system (DCS)), they access
IoT data through APIs, and use simulation and optimization systems namely advanced process control (APC) and
real-time optimization (RTO). The DT is not tested by defining metrics but the prediction model is tested using four
evaluation criteria: the model accuracy ratio (MAR), the root mean square error (RMSE), the variance interpretation
rate (VIR), and Pearson’s correlation coecient (PCC).
Since DT is a twin of the physical asset, implementing it requires the same amount of knowledge, that is required
when creating the physical twin. For any domain, it is essential to have the domain knowledge to understand the
intricacies involved in building and operating the physical asset, such as, how dierent components link together, how
much to weigh parameters during optimisation, etc. Hence, domain experts are essential to implement a DT. [32]
suggests the use of adaptive intelligence for DT which demands a great deal of human expertise, as human knowledge
forms the cardinal element for this artificially intelligent system,.
The above analysis of DT in various sectors indicates how DT is domain-dependent and so are the related chal-
lenges which we discuss ahead.
4.2.1. Similarity &Dierences Across Dierent Domains
To further lay a clear foundation of the concept of DT, we discuss a the similarity and dierences that DTs possess
when being implemented across dierent domains:
Similarity The components of DT mentioned in Section 3.2 should remain the same (though the level of implemen-
tation is subject to domain).
Dierences The following domain dependent questions and implications arise:
1. Particular domain poses greater challenges for the implementation of DT:
Some digital twins might be more feasible and easier to implement than others (due to scale, resources
required and number of components). Same goes for data collection. For instance, the aerospace sector has
numerous components to handle or a supply chain on a global level has many parameters and multiple inventory
points; these cases will prove to be a challenging joint optimisation problem.
2. How realistic the implementation of the DT is with respect to the actual physical asset:
Depending on the complexity of the physical asset, the implementation of the DT will dier from domain
to domain. The gap between the ideal design and practical design might be large for domains like aerospace
and supply chain (w.r.t simulation this is discussed in [80]), and small for PLM and health and prognostics.
3. Evaluation of the implementation of DT:
The evaluation metrics for performance of DT are dependent on the domain. As some domains will lay
more importance on some sub-components such as either of prediction or real-time monitoring, consequently,
the performance evaluation of the DT will be subject to the parameters most important for the specific domain
(such as for aerospace it is mitigation plan).
4. The interoperability issues in the range of dierent softwares being used:
If domains use particular type of software (SCADA, ERP, etc) for sub-tasks of product and logistics man-
agement, how to make the DT software compatible with these systems, which are specific to the domain sectors.
5. The eect of domain knowledge on the implementation of DT:
Domain experts are imperative and essential for designing a DT. A thorough knowledge of the particular
domain and its every sub-component is crucial for designing a DT.
6. What is stopping the spread implementation of DT across a domain:
Is it the lack of liaison between domain experts and DT providers?
Are some domains too complex to implement DT or need a simpler model?
Are the lack of compatibility within the domain specific components concerning IoT, data and machine
Learning a problem?
Is the regulation and standards across domains an issue?
4.3. Industrial Implementations
There are a number of companies now which are investing in DT technology, or providing DT software for clients,
or using DT functionality for themselves:
1. Investing in DT
Signify Philips is exploring the concept of digital twin for lighting [19] by digitising lighting. They claim
it to oer emergency services, real-time monitoring and predictive maintenance.
2. Providing DT as a service
Philips is also providing DT technology for the use of healthcare systems [32] to get early signs of warning
regarding technical issues in Magnetic resonance imaging (MRI), computerized tomography (CT) scan
like medical systems. This could save the downtime that faulty technical systems have on clinical spaces.
IBM is transforming the Port of Rotterdam using Digital Twin [6] for monitoring and eciency. IBM is
also providing a DT software for PLM [34]. Along similar lines, Siemens has a model to implement the
digital twin of power grid in Finland [68], [21] and another for Red Bull in Formula 1 racing [7].
Companies, namely, Dassault System’s 3DEXPERIENCE [71] (which has also built digital twin of Sin-
gapore [61]), AnyLogic [82], Ansys [3], Visualiz [81], PwC [56], Bosch [5], SAP [64] and Azure [51],
[31] provide DT implementation software for clients (though the use of ML in these softwares isn’t clear).
Apart from these, Oracle provides DT simulator as part of its cloud service [55,12]. GE has contributed
to the DT literature and holds 2 patents for DT [15,9] and has a commercial software Predix [24]. The
open source community has also explored the concept of DT and, an open source technology Eclipse Ditto
implements IoT based DT models [17].
3. Using DT for own use
DHL has implemented first digital twin supply chain for Tetra Pak’s warehouse in Asia Pacific in Singa-
pore [16]. Though this is real-time monitoring, the use of machine learning seems unclear and is plausibly
next to none.
BP has employed a surveillance and simulation system called APEX for creating virtual copies of its pro-
duction systems (again ML use unknown) [8].
To the best of our knowledge, none of the above softwares are implemented at global level (but only at country or city
level), and more detailed and concrete analysis on this is subject to further investigation.
4.4. The gap between the ideal DT and practical DT
Depending on the technologies such as IoT, big data and machine learning, there may be a huge gap between
the ideal implementation of DT and the practical one (such as whether the required advancement in technology is
currently available or is subject to further research). Cost and number of available resources could also contribute to
increase in this implementation gap. This discussion is subject to more investigation and availability of the actual DT
software and its initial design model.
From the previous sections, we do not have any information on the level of implementation. Some use outsourced
DT software, but do not describe the frameworks. It is hard to know without evaluation metric-based evidence how
successful the current implementations of DT are. However, most papers have claimed the following to be success-
ful uses of DT — data handling, real-time monitoring, simulation testing and optimisation [73,53]. However, as
mentioned, these implementations without all the components described in Section 3.2 do not fully adhere to the
conceptual definition of DT. Therefore, it is necessary to have a consensus reached regarding a unified, standard
architecture of DT.
4.5. Machine Learning in DT
Conventional knowledge-based methods are based on one-time machine learning output, whereas a digital twin is
a continuous interactive process. Real-time machine learning capacity is what dierentiates a DT from a simulator
or a real-time monitoring tool. Analytics is essential, as one of the cardinal uses of a DT is to be able to reliably and
accurately output how a physical asset would behave in conditions which have not arisen yet, using the real-time data
it is receiving – in other words – for ’testing’ the physical asset in an unforeseen situation. The other cardinal use of
ML in DT is to predict an impending problem which needs attention.
One work that explores the digital twin with machine learning is [53]. It presents a proof-of-concept for using
machine learning with digital twin in the petrochemical industry, by making use of ML algorithms such as random
forest, AdaBoost, XGBoost, gradient boosting decision tree (GBDT), LightGBM, and neural networks. However,
information on the exact implementation methodology and software for DT, and how the entire proof-of-concept was
implemented in real time feedback loop, is absent.
Another work, [18], discusses using a simulation software along with machine learning for a product, and terms
this as ’the digital twin approach’. This is the result of inconsistency in a unified architecture and definition of digital
twin as calling a simulation software with no real-time synchronous connection to the physical product as a digital twin
might be correct according to some definitions while wrong according to other. Hence, it is important for a technology
to have all the components deduced in Section 3.2, to call it a DT and dierentiate it from other technologies.
The ’self-evolving’ nature of DT, where a DT can improve itself from the true results of its predictions, can only
be implemented using machine learning. Machine learning can also help in creating resilient DT. The promising work
[14] uses machine learning in DT not as a feedback mechanism but as an error resolver rather than an analytical tool.
Cronrath et al. [14] use reinforcement learning to make the digital twin resilient to data or model errors, and to learn
to fix such inconsistencies itself.
Despite still being popularly marketed as a DT software by companies like IBM [34], SAP [64], Siemens [68],
the published literature on using machine learning for Digital Twin is scanty and the extent of use of ML by these
companies is uncertain.
There are many advantages of implementing an ML model, however, the following technical challenges are im-
minent to these models, which we suspect might be the reason for lack of ML in the existing DT models:
1. Real-time synchronisation for continuous feedback:
High-fidelity synchronisation between components is required for real-time feedback which depends on
the IoT devices and network connection.
2. Joint optimisation issues:
As described earlier, multiple parameters in the network present a complex multi-objective optimisation
problem. For very large firms, nailing down the sub-optimisation problems to a single one, can be a challenging
task. This requires knowledge of both the domain and machine learning.
3. Challenges related to data required for machine learning (discussed in next section)
4.6. Big Data in DT
DT is used in sectors which have multiple components resulting in multiple parameters. Hence the data collected
from these sources ends up being a large high-dimensional dataset. Moreover, if the time frequencies of the data
collected from these dierent components do not match, the resulting data can be fragmented. Therefore, there
exist time lags in time series data. Additionally, collection of data from multiple inter-connected and not-connected
components, with high-stream synchronisation and integrating this data, is a challenging task in terms of technology,
implementation, cost and resources [53].
Min et al. [53] identify two major issues and sub-issues related to implementing digital twin in real world with its
data component:
1. As constantly changing environment requires a well-thought of tool, concrete and practical frameworks are
needed for big data application to the dynamically changing environment of DT.
2. Data processing issues for time series data: Data gathered from IoT devices in the factory have large dimensions.
Moreover, the data collected may have dierent time cycles.
The authors identify the above problems and solve the problem of using data from dierent time frequencies by
coming up with a method to generate same frequency time series data.
Tao et al. [73] make an attempt to define the steps for data preprocessing, data mining and data optimisation for
DT, which are essential for large datasets. Another work, [72], discusses the limitations of gathering large datasets.
The work also assesses the problems in data management specific to the PLM domain, such as existence of duplicate
data, absence of big data analysis and existence of disintegrated data in dierent phases of PLM. The paper proposes
DT as a solution for comparing this inconsistent data with real values in PLM. Lu et al. [46] points that domain
knowledge could help dealing with missing data issues.
Despite being domain-dependent, there are issues regarding data in DT which exist across domains:
1. Gathering data from dierent IoT devices: Depends on the IoT technology, network connection and intercon-
nectivity among dierent components.
2. Handling large datasets: Requires proper resources for storage and preprocessing steps.
3. High dimension of the datasets: Requires expertise in big data and data reduction techniques.
4. Synchronising data from dierent components frequency-wise: Requires standardised IoT devices, connection
and research.
Qi and Tao [57] compare and contrast Big Data and DT, within the context of smart manufacturing & Industry
4.0. They consider that the main role of DT is the cyber-physical integration, so that DT maintains an accurate
representation of the real system, which could then be used to predict and feedback optimisation decisions to the real
Huang et al. [33] present a blockchain solution to deal with the problem of data integrity in a system. Despite being
promising for health systems, financial systems and cross-industry collaborations, this concept adds to the complexity
of data handling. Adding blockchain technology to DT and optimising it with the rest of the components would be
the next challenge [45] after integrating big data solutions in DT.
5. Industries where DT can be majorly beneficial
Though DT is a technology which benefits any industry in general, there are certain sectors which can benefit
1. Industries where creating physical prototypes is expensive, requires resources and is time-consuming (such as
aerospace, supply chain, manufacturing): Rather than spending time and money for building multiple prototypes
for testing a product, digital twin oers a much more ecient and feasible solution.
2. Industries in which extreme testing is required and performing such tests is hard/not possible in the labs (such
aerospace, PHM): Tests which cannot be performed in the lab can be simulated by the DT.
3. Industries which require real-time monitoring and mitigation plans for dealing with ’emergent behaviour’ [29]
(such as health systems, supply chain): Keeping an eye on the real-time status of the physical asset, and being
alerted through predictions for an imminent problem can be both ecient and eective. This is especially useful
for those organisations which need to make very quick decisions to prevent huge losses.
4. Industries with multiple parameters, which could be optimised jointly (such as manufacturing, supply chain):
For very large organisations, maintaining and monitoring all sub-components can be an extremely dicult task.
Real-time monitoring of all sub-components and joint holistic analytics on such huge models can be beneficial
(this goes along the obstacle of ’siloing’ presented in [29], where due to multiple sub-domains in the system,
information remains fragmented).
Despite being dependent on multiple technologies (explored in detail in [58]) which requires experts and resources,
DT can lead to huge cost reductions for the one time investment [29]. DT can enable shorter design cycles [25], save
cost, resources and time on prototyping [32,63], and predict impending dangers in time to mitigate them. This
cost reduction could possibly be used as a metric of the performance of DT for profit-oriented companies, i.e., the
DT-enabled cost reductions.
6. Current Challenges and Limitations in DT
6.1. Challenges
Current models face the following challenges, some of which are weighed more depending on the domain the DT
is being implemented. These challenges are mainly technical:
1. Data handling (high dimensional [65]; time series, multi-modal and multi-source data communication [72]: In
large organisations the time-series data is collected from numerous IoT devices, resulting in high dimensions
for the dataset. Collecting data from a considerably large number of IoT devices, collating it according to time
frequencies and preprocessing it for input to machine learning is a challenging big data task.
2. High-fidelity 2 way synchronisation is especially hard for large-scale industries, requires resources and high-
stream IoT connection [27,65,72].
3. Simulation software and simulation-based optimisation: Building a software for DT demands a team of pro-
grammers, developers and domain experts to test the suitability of the software for the particular task. Moreover,
simulation-based optimisation provides faster and ecient solutions [2]. Conventionally, solutions calculated
by analytical models are fed to the simulation software manually rather than optimisation being done on the
simulation software by use of mechanisms like reinforcement learning.
4. Joint optimisation for all the sub-problems: Depending on the size of the industry, the number of parameters
involved in the machine learning optimization problem can be huge, as each industry has smaller optimization
problems which are solved at various sub-levels. Combining these sub-problems into one optimization problem
and choosing the right machine learning tools is a challenging task. True predictions of the entire physical asset
can only be made if all the components of the asset are taken into consideration, resulting in a joint machine
learning-based optimisation problem. This kind of unified ML model will easily facilitate and weigh priorities
in a system, and hence perform better analytics.
5. Global Implementations (currently only city level or country level implementations exists): Current implemen-
tations (as discussed in section 4.3) are mainly focused on city/country. For larger systems, such as supply chain
networks and logistics, etc, global implementation level is more desirable.
6. Deep learning with DT: Implementing deep learning solutions requires computational resources, expertise and
research [42]. Managing high dimensional data, with the various other software used by an industry and com-
bining these with expert deep learning skills and equipment is a tedious task. New methods like continual
learning [43] and federated learning [85] seem promising for the use of DT but require further research.
7. Interoperability with existing softwares being used in a production lifecyle [30]: Industries use multiple soft-
ware for inventory, product management, etc. The compatibility of DT with these is a challenging issue, tackling
which might lead to delay in implementations.
8. Cybersecurity concerns, IoT security, cross industrial partners security [30]: With the digital twin operating
across multiple industrial partners and inventory sites, the security concerns are inevitable. Not only the cross
industry security concerns but also the leak of real-time monitoring data can be hazardous to a firm.
6.2. Limitations
Despite the advantages that DT brings to an industry, depending on various factors, there exist certain limitations:
1. Cost: Since implementing DT and profiting from it is a timely process, DT can be costly if the life and span of
a project is small.
2. Complexity: As DT requires interoperability among various components, real-time tools, formulating a joint
optimisation problem and big data resources, putting these together can be time-consuming for an industry and
may lead to unwanted distractions.
3. Research Add-on: Like any technology, DT will be needed to be updated according to the recent developments
in the technologies it relies on (IoT, big data, machine learning). Industries with long-term DT use will therefore
need to continuously invest in this research, which might lead to added cost.
7. Open Research Questions and Future Steps
This review has posed and answered some new key questions in the field. Nevertheless, there are still some
questions which do not have a definite answer yet:
1. How to quantify the performance of a DT: Quantitative metrics are essential to understand the real impact of
using a digital twin. It can also help in knowing how the performance of a digital twin can be improved. For a
complete evaluation of the performance of DT and its suitability to the domain concerned, these error metrics
have to be categorised as domain-dependent and domain-independent (such as [69] proposed error metrics
specific to an aircraft model). This could also include uncertainty quantification to calculate the confidence
levels of outputs of DT. Also, the self-evolving nature of DT can be implemented only if the DT can assess its
own performance quantitatively.
2. How to rectify data and model errors: Information shared across devices might contain noise and not always be
accurate. Incorrect information models can be detrimental to the entire DT system. If the data or model fed to
the DT is inaccurate, it can result in huge losses, due to incorrect predictions or test results (the human in the
end of the ’loop’ plays an important role here in deciding whether to follow the DT’s recommendation or not)
To build resilient and self-evolving DT, more research is needed. Machine learning is a promising direction for
this [14]. These could include, anomaly detection (DT could check whether a particular data is misfed), and
traceability (logs of the data used by DT to generate outputs could be available to check how DT generated an
output. This could also help in improving the DT and for feature engineering in machine learning). No error
quantification metric: It is essential to quantify the error made by the DT in order to understand its performance.
It also helps the managers in deciding how much to rely on the outcome of the DT, depending on its confidence
levels. Without error quantification, to follow or not to follow the feedback from DT is taking a chance.
Apart from tackling these issues, we believe the following future steps will enhance the advancement of DT and
lead to its widespread useful implementation:
1. A Formal Definition of DT: As a lack of consensus exists for the definition of digital twin, having a formal
definition will help to clarify the concept and reach a universally accepted concept.
2. IoT standards required for DT: Since DT relies heavily on IoT devices for capturing and sharing data, real-time
synchronisation and monitoring, knowing what IoT standards are best-suited for these operations will enhance
the acceptance of DT and make it easier for widespread adoption. Lu et al. [46] recognises the same need for a
data communication standard.
3. Regulations at enterprise and global levels [72]: As many companies collaborate across industries, having
proper legal-binding regulations on the data used in DT is crucial for smooth operation. This also applies to
sites spread across the globe which need to adhere to laws applicable of the particular country.
4. Liaising with domain experts to spread DT across sectors: Communicating with domain experts will facilitate
easier implementation and acceptance of DT into new domains. As domain experts know all there is to know
about domain-specific requirements, they can be very useful in designing the DT. Once a design is ready, the
implementation of the DT framework can be managed by programmers and developers.
8. Conclusions
There have been many papers defining the digital twin and its components, however, no universal model exists till
date which is adopted widely. We reviewed the past work and defined our DT reference model, based on these past
works which include important components. These components define the essence of a digital twin, such as the bi-
jective relation between DT and the physical asset. These components also highlight how to evaluate the performance
and integrity of a DT system (testing and security components).
As a tool which can have many sub-components spread across collaborators and industry partners, developing reg-
ulations and security mechanisms is imperative for widespread of adoption of DT to overcome the concerns regarding
data sharing. Machine learning techniques like federated learning are promising in this direction too.
Digital twin is a powerful tool with great capabilities combining simulation, autonomy, agent-based modeling,
machine learning, prototyping, optimisation and big data into one. Theses sub-systems can be tailored and prioritised
depending on the application domain needs. However, the advancement and research in these subsystems at times
create a hindrance for the development of DT. We conclude that, as up-and-coming the DT technology is, there are
certain technical and domain-dependent challenges that still need to be addressed. The technology does depend on its
counterparts of IoT, machine learning and data, however, a seamless integration of all these leads to the powerful and
ecient product that a DT is.
The work reported here was sponsored by Research England’s Connecting Capability Fund award CCF18-7157
- Promoting the Internet of Things via Collaboration between HEIs and Industry (Pitch-In). We thank Kate Price
Thomas for her suggestions for the literature review and industrial works.
[1] Abramovici, M., G ¨
obel, J.C., Dang, H.B., 2016. Semantic data management for the development and continuous reconfiguration of smart
products and systems. CIRP Annals - Manufacturing Technology 65, 185–188. doi:10.1016/j.cirp.2016.04.051.
[2] Alrabghi, A., Tiwari, A., 2015. State of the art in simulation-based optimisation for maintenance systems. doi:10.1016/j.cie.2014.12.
[3] Ansys, 2020. Digital Twin Physics-Based Simulation & Analytics — Ansys. URL:
[4] Ballot, E., Montreuil, B., Meller, R., 2014. The Physical Internet URL: https://hal-mines-
[5] Bosch, . Bosch IoT Things: Managed inventory of digital twins - Developer Bosch IoT Suite. URL: https://developer.
[6] Boyles, R., 2019. How the Port of Rotterdam is using IBM digital twin technology to transform itself from the biggest to the smartest. -
Internet of Things blog. URL:
[7] Breuer, H., 2018. The Future of Manufacturing — Digital Twin — Siemens Global Website. URL:
en/company/stories/research-technologies/digitaltwin/the- future-of-manufacturing- plm-red-bull- racing.
[8] British Petroleum, 2018. Twin Win for Oil and Gas Production. URL:
news-and- insights/bp-magazine/apex-digital- system.html.
[9] Canedo, Z.S.A.M., . Digital twins for energy ecient asset maintenance. Technical Report.
[10] Cearley, D., 2019. Top 10 Strategic Technology Trends for 2019. URL:
top-10- strategic-technology-trends- for-2019-digital- twin.
[11] Cearley, D.W., Velosa, A., Kerremans, M., Walker, M.J., Burke, B., 2018. Top 10 Strategic Technology Trends for 2018: Digital Twins. URL: strategic-technology-trends- for-2018-digital- twin0.
[12] Corporation, O., 2020. About the Oracle IoT Digital Twin Implementation .
[13] Cortada, J.W., 2018. Information and the Modern Corporation. doi:10.7551/mitpress/9054.001.0001.
[14] Cronrath, C., Aderiani, A.R., Lennartson, B., 2019. Enhancing digital twins through reinforcement learning. IEEE International Conference
on Automation Science and Engineering 2019-Augus, 293–298. doi:10.1109/COASE.2019.8842888.
[15] Dell ’ Anno, M.J., . Digital twin of twinned physical system. Technical Report.
[16] DHL, 2019. DHL Supply Chain Partners Tetra Pak to Implement Its First Digital Twin Warehouse in Asia
Pacific — DHL — Indonesia. URL: archive/2019/
dhl-supply- chain-partners-tetra- pak-to-implement- its-first-digital- twin-warehouse-in- asia-pacific.html.
[17] Ditto, E., . Eclipse Ditto open source framework for digital twins in the IoT. URL:
[18] Dr¨
oder, K., Bobka, P., Germann, T., Gabriel, F., Dietrich, F., 2018. A machine learning-enhanced digital twin approach for human-robot-
collaboration, in: Procedia CIRP, Elsevier B.V.. pp. 187–192. doi:10.1016/j.procir.2018.02.010.
[19] Duine, P., . Digital twins: what do they mean for lighting? — Philips lighting. URL:
oem-emea/stay- connected/digital-twins.
[20] Enders, M.R., Hoßbach, N., 2019. Dimensions of Digital Twin Applications -A Literature Review Dimensions of Digital Twin Applications-A
Literature Review Completed Research. Technical Report.
[21] Engelhardt, M., 2018. Digital Twin Fingrid — Energy — United Kingdom. URL:
stories/energy/digital-twin- fingrid.html.
[22] Forni, A.A., 2016. Gartner Identifies the Top 10 Strategic Technology Trends for 2017. URL:
press-releases/2016- 10-18-gartner- identifies-the-top- 10-strategic-technology- trends-for-2017.
[23] Gabor, T., Belzner, L., Kiermeier, M., Beck, M.T., Neitz, A., 2016. A simulation-based architecture for smart cyber-physical systems, in:
2016 IEEE International Conference on Autonomic Computing (ICAC), IEEE. pp. 374–379.
[24] GE, 2018. Digital Twin — GE Digital. URL:
[25] Gery CTO, E., . Industry Transformation with IBM Digital Twin. Technical Report.
[26] Glaessgen, E.H., Stargel, D.S., 2012. The Digital Twin Paradigm for Future NASA and U.S. Air Force Vehicles. Technical Report. URL:
[27] Grieves, M., 2014. Digital twin: manufacturing excellence through virtual factory replication. White paper 1, 1–7. URL: http:
//{_}Dr{_}Grieves{_}DigitalTwin{_}ManufacturingExcellence.php, doi:10.5281/
[28] Grieves, M., 2015. Digital Twin : Manufacturing Excellence through Virtual Factory Replication This paper intro-
duces the concept of a A Whitepaper by Dr . Michael Grieves. White Paper URL:
[29] Grieves, M., Vickers, J., 2017. Digital twin: Mitigating unpredictable, undesirable emergent behavior in complex systems, in: Transdisci-
plinary perspectives on complex systems. Springer, pp. 85–113.
[30] Grieves, M.W., 2019. Virtually intelligent product systems: Digital and physical twins. Complex Systems Engineering: Theory and Practice
, 175–200.
[31] Hoof, B.V., 2018. Announcing Azure Digital Twins: Create digital replicas of spaces and infrastructure using cloud,
AI and IoT — Azure blog and updates — Microsoft Azure. URL: gb/blog/
announcing-azure- digital-twins-create- digital-replicas-of- spaces-and-infrastructure- using-cloud-ai- and-iot/.
[32] van Houten, H., 2020. The rise of the digital twin: how healthcare can benefit - Blog —
Philips. URL: matters/
20180830-the- rise-of-the- digital-twin-how- healthcare-can-benefit.html.
[33] Huang, S., Wang, G., Yan, Y., Fang, X., 2020. Blockchain-based data management for digital twin of product. Journal of Manufacturing
Systems 54, 361–371. URL:, doi:10.1016/j.jmsy.2020.01.009.
[34] IBM, . IBM Digital Twin Exchange - Overview - United States. URL:
digital-twin- exchange.
[35] Ivanov, D., Vladimirovich Sokolov, B., Dolgui, A., Das, A., Sokolov, B., 2019. Digital Supply Chain Twins: Managing the Ripple Eect,
Resilience, and Disruption Risks by Data-Driven Optimization, Simulation, and Visibility International Journal of Integrated Supply Man-
agement View project Structural Dynamics and Resilience in Suppl URL:,
doi:10.1007/978-3- 030-14302-2_15.
[36] Jazdi, N., 2014. Cyber physical systems in the context of Industry 4.0, in: Proceedings of 2014 IEEE International Conference on Automation,
Quality and Testing, Robotics, AQTR 2014, IEEE Computer Society. doi:10.1109/AQTR.2014.6857843.
[37] Jeschke, S., Brecher, C., Meisen, T., ¨
Ozdemir, D., Eschert, T., 2017. Industrial Internet of Things and Cyber Manufacturing Sys-
tems, Springer, Cham, pp. 3–19. URL: 42559-7{_}1, doi:10.1007/
978-3- 319-42559-7_1.
[38] Kiritsis, D., 2011. Closed-loop PLM for intelligent products in the era of the Internet of things. CAD Computer Aided Design 43, 479–501.
[39] 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. doi:10.1016/j.ifacol.2018.08.474.
[40] Lee, J., Lapira, E., Bagheri, B., an Kao, H., 2013. Recent advances and trends in predictive manufacturing systems in big data environment.
Manufacturing Letters 1, 38–41. doi:10.1016/j.mfglet.2013.09.005.
[41] Li, L., 2018. China’s manufacturing locus in 2025: With a comparison of “Made-in-China 2025” and “Industry 4.0”. Technological
Forecasting and Social Change 135, 66–74. doi:10.1016/j.techfore.2017.05.028.
[42] Liu, Q., Zhang, H., Leng, J., Chen, X., 2018. International Journal of Production Research Digital twin-driven rapid individualised designing
of automated flow-shop manufacturing system Digital twin-driven rapid individualised designing of automated flow-shop manufacturing sys-
tem URL:, doi:10.1080/00207543.2018.
[43] Lopez-Paz, D., Ranzato, M..A., . Gradient Episodic Memory for Continual Learning. Technical Report. URL:
[44] Lu, Y., 2017. Industry 4.0: A survey on technologies, applications and open research issues. doi:10.1016/j.jii.2017.04.005.
[45] Lu, Y., 2019. The blockchain: State-of-the-art and research challenges. doi:10.1016/j.jii.2019.04.002.
[46] 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. Robotics and Computer-Integrated Manufacturing 61, 101837.
[47], 2020. Digital Twin Market by Source, Type, Application & Geography— COVID-19 Impact Analy-
sis—MarketsandMarkets. URL:
[48] McFarlane, D., Sarma, S., Chirn, J.L., Wong, C.Y., Ashton, K., 2002. The intelligent product in manufacturing control and management.
Technical Report 1.
[49] McFarlane, D.C., Bussmann, S., 2000. Developments in holonic production planning and control. Production Planning and Control 11,
522–536. doi:10.1080/095372800414089.
[50] Meyer, G.G., Fr¨
amling, K., Holmstr¨
om, J., 2008. Intelligent Products: a survey. Technical Report.
[51] Microsoft, 2018. Azure Digital Twins — Microsoft Azure. URL:
[52] Miller, A.M., Alvarez, R., Hartman, N., 2018. Towards an extended model-based definition for the digital twin 15, 880–891. URL: https:
//, doi:10.1080/16864360.2018.1462569.
[53] Min, Q., Lu, Y., Liu, Z., Su, C., Wang, B., 2019. Machine Learning based Digital Twin Framework for Production Optimization in Petro-
chemical Industry. International Journal of Information Management 49, 502–519. doi:10.1016/j.ijinfomgt.2019.05.020.
[54] Negri, E., Fumagalli, L., Macchi, M., 2017. A Review of the Roles of Digital Twin in CPS-based Production Systems. Procedia Manufacturing
11, 939–948. doi:10.1016/j.promfg.2017.07.198.
[55] Oracle, 2017. Digital Twins for IoT Applications: A Comprehensive Approach to Implementing IoT Digital Twins. Technical Report January.
URL: for-iot-apps- wp-3491953.pdf.
[56] PwC, 2020. Digital Twin: PwC. URL:
[57] Qi, Q., Tao, F., 2018. Digital twin and big data towards smart manufacturing and industry 4.0: 360 degree comparison. Ieee Access 6,
[58] Qi, Q., Tao, F., Hu, T., Anwer, N., Liu, A., Wei, Y., Wang, L., Nee, A., 2019. Enabling technologies and tools for digital twin. Journal of
Manufacturing Systems .
[59] 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. doi:10.2514/6.2013- 1578.
[60] R´
ıos, J., Hern´
andez, J.C., Oliva, M., Mas, F., 2015. Product avatar as digital counterpart of a physical individual product: Literature review
and implications in an aircraft. Advances in Transdisciplinary Engineering 2, 657–666. doi:10.3233/978- 1-61499-544- 9-657.
[61] Rocker, I., 2015. Virtual Singapore and the Economy of the Digital Twin. URL:
virtual-singapore- and-the-economy- of-the-digital- twin/.
[62] Rosen, R., Von Wichert, G., Lo, G., Bettenhausen, K.D., 2015. About the importance of autonomy and digital twins for the future of
manufacturing, in: IFAC-PapersOnLine, pp. 567–572. doi:10.1016/j.ifacol.2015.06.141.
[63] Roy, R., Stark, R., Tracht, K., Takata, S., Mori, M., 2016. Continuous maintenance and the future – Foundations and technological challenges.
CIRP Annals - Manufacturing Technology 65, 667–688. doi:10.1016/j.cirp.2016.06.006.
[64] SAP, . SAP Digital Twin Software & Technology. URL: supply-chain/
[65] Schleich, B., Anwer, N., Mathieu, L., Wartzack, S., 2017. Shaping the digital twin for design and production engineering Shaping the digital
twin for de-sign and production engineering. CIRP Annals-Manufacturing Technology 66. URL: https://hal.archives- ouvertes.
fr/hal-01513846, doi:10.1016/j.cirp.2017.04.040.
[66] Schroeder, G.N., Steinmetz, C., Pereira, C.E., Espindola, D.B., 2016. Digital Twin Data Modeling with AutomationML and a Communication
Methodology for Data Exchange. IFAC-PapersOnLine 49, 12–17. doi:10.1016/j.ifacol.2016.11.115.
[67] Shafto, M., Rich, M.C., Glaessgen, D.E., Kemp, C., Lemoigne, J., Wang, L., 2010. DRAFT MoDeling, SiMulATion, inFoRMATion Tech-
nology & PRoceSSing RoADMAP Technology Area 11. Technical Report.
[68] Siemens AG, 2017. Factsheet: For a digital twin of the grid - Siemens solution enables a single digital grid model of the Finnish power
system. Technical Report.
[69] Smarslok, B.P., Culler, A.J., Mahadevan, S., 2013. Error Quantification And Confidence Assessment Of Aerothermal Model Predictions
For Hypersonic Aircraft (Preprint) Hypersonic Sciences Branch High Speed Systems Division Stinfo Copy Air Force Research Laboratory
Aerospace Systems Directorate Wright-patter. Technical Report. URL:
[70] Sternberg, H., Norrman, A., 2017. The Physical Internet – review, analysis and future research agenda. doi:10.1108/
IJPDLM-12- 2016-0353.
[71] Syst`
emes, D., 2020. 3DEXPERIENCE®New Releases - Dassault Syst`
emes®. URL: services/
[72] Tao, F., Cheng, J., Qi, Q., Zhang, M., Zhang, H., Sui, F., 2018. Digital twin-driven product design, manufacturing and service with big data.
The International Journal of Advanced Manufacturing Technology 94, 3563–3576. doi:10.1007/s00170-017-0233- 1.
[73] Tao, F., Zhang, H., Liu, A., Nee, A.Y., 2019. Digital Twin in Industry: State-of-the-Art. IEEE Transactions on Industrial Informatics 15,
2405–2415. doi:10.1109/TII.2018.2873186.
[74] Terkaj, W., Pedrielli, G., Sacco, M., 2012. Virtual factory data model, in: Proceedings of the workshop on ontology and semantic web for
manufacturing, Graz, Austria, pp. 29–43.
[75] Terkaj, W., Urgo, M., 2014. Ontology-based modeling of production systems for design and performance evaluation, in: Proceedings -
2014 12th IEEE International Conference on Industrial Informatics, INDIN 2014, Institute of Electrical and Electronics Engineers Inc.. pp.
748–753. doi:10.1109/INDIN.2014.6945606.
[76] Tuegel, E., 2012. The Airframe Digital Twin: Some Challenges to Realization, in: 53rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural
Dynamics and Materials Conference<BR>20th AIAA/ASME/AHS Adaptive Structures Conference<BR>14th AIAA, American Institute
of Aeronautics and Astronautics, Reston, Virigina. URL:, doi:10.2514/6.
[77] Tuegel, E.J., Ingraea, A.R., Eason, T.G., Spottswood, S.M., 2011. Reengineering Aircraft Structural Life Prediction Using a Digital Twin.
International Journal of Aerospace Engineering 2011, 14. doi:10.1155/2011/154798.
[78] Valckenaers, P., Germain, B.S., Verstraete, P., Van Belle, J., Hadeli, Van Brussel, H., 2009. Intelligent products: Agere versus Essere.
Computers in Industry 60, 217–228. doi:10.1016/j.compind.2008.12.008.
[79] Valckenaers, P., Hadeli, H., Germain, B.S., Verstraete, P., Van Belle, J., Van Brussel, H., 2007. From intelligent agents to intelligent beings,
in: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics),
Springer, Berlin, Heidelberg. pp. 17–26. doi:10.1007/978-3-540-74481- 8_3.
[80] Vieira, A.A., Dias, L.M., Santos, M.Y., Pereira, G.A., Oliveira, J.A., 2020. Supply chain data integration: A literature review. Journal of
Industrial Information Integration 19, 100161. URL:, doi:10.1016/j.jii.2020.
[81] Visualiz, 2020. Data visualization for VR, big data, BI and IoT - VISUALIZ. URL:
[82] Wilkinson, G., 2019. Blog – digital twin – AnyLogic Simulation Software. URL:
[83] Wong, C.Y., McFarlane, D., Zaharudin, A.A., Agarwal, V., 2002. The intelligent product driven supply chain. Technical Report. doi:10.
[84] Xu, X., 2012. From cloud computing to cloud manufacturing. Robotics and Computer-Integrated Manufacturing 28, 75–86. doi:10.1016/
[85] Yang, Q., Liu, Y., Chen, T., Tong, Y., 2019. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems
and Technology 10, 1–19. URL:, doi:10.1145/3298981,arXiv:1902.04885.
ResearchGate has not been able to resolve any citations for this publication.
Conference Paper
Full-text available
The use of Digital Twins has gained attraction in research and practice in recent years. Digital Twins are virtual representations of physical objects and they can be connected with their physical counterparts. Through this connection, Digital Twins contribute to the convergence of the real and the virtual world. While existing literature reviews focus strongly on the manufacturing industry, this paper analyzes Digital Twin applications across industries. Based on a systematic literature review, this paper examines 87 Digital Twin applications and proposes a classification scheme with six dimensions to describe the applications identified. The concept of Digital Twins is currently still underrepresented in Information Systems research, which opens up further research opportunities.
Full-text available
The concept of digital twin was originally proposed to assist product lifecycle management through the high-fidelity virtual product. In product lifecycle, there are many participators constructing a complicated network with enormous product lifecycle data. It is difficult to manage data of digital twin efficiently and securely from the perspectives of data storage, data access, data sharing, and data authenticity among the complicated network. Moreover, the virtual product is always updated to the latest state of the physical product by overwriting the history virtual product that records the development process of digital twin. To address abovementioned data management problems simultaneously, a data management method for digital twin of product based on blockchain technology is proposed in this paper. A peer-to-peer network is constructed to enhance data sharing efficiency among participators. The basic structure of blockchain is introduced, including block, chain, and transaction. The blockchain is used for data storage through cryptography that ensuring only eligible participators can access the corresponding data. The change sensitive characteristic of blockchain can ensure data authenticity as well. In addition, the concept of smart contract can be used to execute some actions automatically to increase data sharing efficiency. A case study is presented to demonstrate the effectiveness of the proposed data management method. The results show that the proposed method can solve the abovementioned data management problems simultaneously.
Full-text available
Digital twin is revolutionizing industry. Fired by sensor updates and history data, the sophisticated models can mirror almost every facet of a product, process or service. In the future, everything in the physical world would be replicated in the digital space through digital twin technology. As a cutting-edge technology, digital twin has received a lot of attention. However, digital twin is far from realizing their potential, which is a complex system and long-drawn process. Researchers must model all the different parts of the objects or systems. Varied types of data needed to be collected and merged. Many researchers and participators in engineering are not clear which technologies and tools should be used. 5-dimension digital twin model provides reference guidance for understanding and implementing digital twin. From the perspective of 5-dimension digital twin model, this paper tries to investigate and summarize the frequently-used enabling technologies and tools for digital twin to provide technologies and tools references for the applications of digital twin in the future.
Full-text available
The idea that the information about a physical object can be separated from the object itself and then mirror or twin that object is a concept referred to as the Digital Twin. The Digital Twin is receiving a great deal of interest from manufacturers who make advanced products that have all the characteristics of complex systems. While the Digital Twin concept is becoming better fleshed out and understood, there is much more work to be accomplished. Specifically, the characteristics of the physical product as these become smart, connected product system (SCPS or Physical Twin) need to be defined and described. The success of the Digital Twin model will rest on the value it creates for both the manufacturers and the users of their products. There will also be new issues, security among the most important, that need to be surfaced and addressed.
In this study, we have compared Germany's “Industry 4.0” and China's “Made-in-China 2025” and estimated China's locus in “Made-in-China 2025”. “Made-in-China 2025” has clear goals, measures and sector focus. Its guiding principles are to enhance industrial capability through innovation-driven manufacturing, optimize the structure of Chinese industry, emphasize quality over quantity, train and attract talent, and achieve green manufacturing and environment. Data show that currently China is no longer the lowest–cost labor market; it is being squeezed by newly emerging low-cost producers such as Vietnam, Cambodia, and Laos. Meanwhile, China is not the strongest player in the high-tech arena; well-established industrialized nations, the US, Germany, and Japan, have all effectively deployed digital technology to create new industrial environments, produce new products, and improve their well-established brands. Having analyzed the data from the World Bank and China's National Bureau of Statistics, we find an upward trajectory in China in manufacturing capability development, research and development commitment, and human capital investment. However, implementing an ambitious strategic plan such as “Made-in-China 2025” is coupled with challenges. This research helps us understand the relationship between technological entrepreneurship and socio-economic changes in emerging economies such as China. Furthermore, the experience accumulated in China can be referenced by both emerging economies and developed nations to advance their technological entrepreneurship.
Digital twin (DT) is one of the most promising enabling technologies for realizing smart manufacturing and Industry 4.0. DTs are characterized by the seamless integration between the cyber and physical spaces. The importance of DTs is increasingly recognized by both academia and industry. It has been almost 15 years since the concept of the DT was initially proposed. To date, many DT applications have been successfully implemented in different industries, including product design, production, prognostics and health management, and some other fields. However, at present, no paper has focused on the review of DT applications in industry. In an effort to understand the development and application of DTs in industry, this paper thoroughly reviews the state-of-the-art of the DT research concerning the key components of DTs, the current development of DTs, and the major DT applications in industry. This paper also outlines the current challenges and some possible directions for future work.
Supply chains (SCs) are dynamic networks subject to uncertainties and risks that may occur anywhere, anytime, and whose consequences affect the entities that comprise such SC, possibly affecting others. In fact, there are several examples wherein the occurrence of certain events resulted in considerable costs. Thus, it is important to ensure that SCs can apply preemptive measures, rather than just react to disruptions that may occur. Simulation tools may play an important role in achieving this, as these tools may be used to test alternative scenarios, as well as to quantify the impact of risks. To fully exploit this possibility, simulation tools should be used as data integration tools, so that the aforementioned analysis can be conducted using data from several relevant sources, thereby improving the quality of such analysis. In this regard, this paper proposes a Systematic Literature Review (SLR) of simulation methods that deal with risks in SCs, with particular emphasis on the type of data integration employed by such works. The obtained results show that researchers tend to simplify the problem at hand, without modeling their entire complexity, and failing to properly integrate data from the involved processes. The analyzed works’ compliance with Industry 4.0 (I4.0) revealed similar conclusions, as it was found that studies tend to disregard some of the main features of simulation in I4.0. In light of the obtained findings, literature gaps are identified, and future research directions are proposed.
Conference Paper
Digital Twins are core enablers of smart and autonomous manufacturing systems. Although they strive to represent their physical counterpart as accurately as possible, slight model or data errors will remain. We present an algorithm to compensate for those residual errors through Reinforcement Learning (RL) and data fed back from the manufacturing system. When learning, the Digital Twin acts as teacher and safety policy to ensure minimal performance. We test the algorithm in a sheet metal assembly context, in which locators of the fixture are optimally adjusted for individual assemblies. Our results show a fast adaption and improved performance of the autonomous system
Digital twins, along with the internet of things (IoT), data mining, and machine learning technologies, offer great potential in the transformation of today’s manufacturing paradigm toward intelligent manufacturing. Production control in petrochemical industry involves complex circumstances and a high demand for timeliness; therefore, agile and smart controls are important components of intelligent manufacturing in the petrochemical industry. This paper proposes a framework and approaches for constructing a digital twin based on the petrochemical industrial IoT, machine learning and a practice loop for information exchange between the physical factory and a virtual digital twin model to realize production control optimization. Unlike traditional production control approaches, this novel approach integrates machine learning and real-time industrial big data to train and optimize digital twin models. It can support petrochemical and other process manufacturing industries to dynamically adapt to the changing environment, respond in a timely manner to changes in the market due to production optimization, and improve economic benefits. Accounting for environmental characteristics, this paper provides concrete solutions for machine learning difficulties in the petrochemical industry, e.g., high data dimensions, time lags and alignment between time series data, and high demand for immediacy. The approaches were evaluated by applying them in the production unit of a petrochemical factory, and a model was trained via industrial IoT data and used to realize intelligent production control based on real-time data. A case study shows the effectiveness of this approach in the petrochemical industry.