nature computational science

Perspective https://doi.org/10.1038/s43588-024-00603-w

Advancements and challenges of digital

twins in industry

Received: 30 July 2023 Fei Tao 1,2,3,4

, He Zhang 2,4
& Chenyuan Zhang2

Accepted: 6 February 2024

Published online: 26 March 2024 Digital twins, which are considered an effective approach to realize the
fusion between virtual and physical spaces, have attracted a substantial
Check for updates
amount of attention in the past decade. With their rapid development in
recent years, digital twins have been applied in various fields, particularly
in industry. However, there are still some gaps to be filled and some
limitations to be addressed. Here we provide a brief overview of digital
twin advancements in industry and highlight the main pitfalls to avoid
and challenges to overcome, to improve the maturity of digital twins and
facilitate large-scale industrial applications in the future.

With the continuous development of information technology in recent for every context. In addition, there are still some misunderstandings
years (such as network communication technology, sensor technology and bottlenecks that hinder further development of digital twins in
and intelligent-embedded technology), the virtual space has had a the industrial field.
much closer relationship with the physical space than ever before. In In this Perspective, we give an overview of digital twin advance-
this context, digital twins, which are often regarded as living in silico ments in industry, followed by a discussion on the pitfalls, challenges
representations of their corresponding physical targets, have grown and promising research avenues based on the current research status
in popularity. Digital twins can iterate and evolve through the seamless and our own findings from practical applications.
connection and fusion between the virtual and physical spaces1, and this
consistency and synchronization can bring many benefits to numer- State of the art in digital twins for industrial
ous services, including real-time monitoring, dynamic optimization applications
and accurate prediction2–4. The value of digital twins has been recog- So far, digital twins have addressed many problems in some indus-
nized by researchers, enterprises and other stakeholders in various trial application scenarios. Here we give an overview of digital twins in
fields, including healthcare5,6, agriculture7, urban science8, aerospace industry from the perspective of theory, technology and applications
engineering9, marine engineering10,11 and even Earth systems12. to better showcase the most recent advancements, understand how
Notably, research and application of digital twins in industry are digital twins are applied and unravel their value.
increasingly widespread, as digital twins align with the contemporary
demands of smart manufacturing and Industry 4.013. At present, more Theory and technology
than 1,000 institutions, including universities, enterprises and research A number of conceptual models of digital twins have been recently
institutes, have carried out related explorations and practice14, which proposed to illustrate their components and connotations, as shown
has already covered the entire life cycle of products and manufactur- in Fig. 1. Figure 1a depicts the three-dimensional digital twin model19,
ing systems, such as realizing rapid engineering design15, real-time which is one of the most popular conceptual models. Figure 1b is as a
monitoring in the operation process16, intelligent decision-making more detailed model that takes into account the product, production
towards continuous changing manufacturing systems17, prognostics and performance. Finally, Fig. 1c describes the overall architecture of
and health management18, and so forth. Different application objects the digital twin for both products and manufacturing systems, which
(such as humans, machines and materials), processing technologies adds data and services to the three-dimensional digital twin model20.
and user requirements directly influence the implementation of digital In addition, physical processes and virtual processes are executed
twins, and as a consequence, digital twins should take into account alternately during the entire life cycle, as shown in the outer ring of
these individual differences of use cases and not always be the same Fig. 1c. The twin model is utilized to support virtual simulations and

Digital Twin International Research Center, International Research Institute for Multidisciplinary Science, Beihang University, Beijing, China. 2School of
1

Automation Science and Electrical Engineering, Beihang University, Beijing, China. 3State Key Laboratory of Virtual Reality Technology and Systems,
Beihang University, Beijing, China. 4Tianmushan Laboratory, Hangzhou, China. e-mail: [email protected]

Nature Computational Science | Volume 4 | March 2024 | 169–177 169

Perspective https://doi.org/10.1038/s43588-024-00603-w

a b
Insights from performance
Digital twin
Physical space Virtual space
Digital twin performance
Physical Digital twin
Virtual
product production
product Information product

Virtual Virtual Real Real

Data product production production product
Verification Validation
Physical Virtual
shop floor shop floor Specification Commissioning Ideal delivery

Automation

Continuous improvement

c Virtual Virtual analysis and

operation decision-making

Physical
operation Physical
execution

Services
Twin
interaction
Virtual
Physical end-of-life
delivery Twin data
Twin Twin
interaction interaction

Iterative Twin Iterative Physical

optimization interaction optimization end-of-life
Virtual
delivery

Physical Virtual
product product

Physical Iterative Virtual Virtual

Physical shop floor shop floor product design
optimization
testing

Physical object Twin model

Virtual Virtual
testing verification
Twin interaction

Twin model-based
virtual process

Physical Physical object-based

Virtual
production real process
optimization
Virtual
production

Fig. 1 | Digital twin conceptual models. a, The three-dimensional digital twin to improve product, production and performance. c, The five-dimensional
model is one of the earliest digital twin conceptual models, which includes digital twin model consists of five dimensions, including physical object, twin
physical object, virtual model, and the interaction between physical and virtual model, services, twin data and twin interaction among them.
spaces. b, The Siemens digital twin model demonstrates how digital twins work

Nature Computational Science | Volume 4 | March 2024 | 169–177 170

Perspective https://doi.org/10.1038/s43588-024-00603-w

optimizations, thereby achieving product design, production, testing To meet practical requirements, such as real-time monitor-
and optimization at a lower cost and higher efficiency. In addition, the ing and dynamic scheduling, multiple models, such as geometric
twin model is continuously enriched, modified and updated based on models, mechanical models, material models and discrete-event
physical information and twin interactions. While these conceptual models, have been constructed40–43. For example, Söderberg et al. pro-
models differ in terms of framework, they all aim to utilize ultra-fidelity posed a digital twin model consisting of the geometric representation
models, rich data and real-time interaction to provide some services of the assembly, kinematic relationships, finite-element-analysis func-
for specific objects. tions, Monte Carlo simulation and material properties, thereby support-
In addition to the conceptual models, reference frameworks ing the geometry assurance in the design and manufacturing stages44.
have also been recently proposed to provide more detailed guide-
lines or templates for implementing digital twins toward certain Twin data computation
needs. For example, to realize autonomy and real-time supervi- Twin data contain both real-time and historical data on various
sion, a manufacturing execution system (MES)-integrated digital elements within manufacturing systems, from both physical and
twin framework was proposed. The framework contains a decision- virtual spaces. At present, more research is devoted to presenting
making level, an MES integration level, and an MES level to deal advanced algorithms or selecting appropriate algorithms oriented to
with error states management during production and reactive dis- specific requirements, such as the Q-learning algorithm for dynamic
assembly during assembly of products21. The field-synchronized scheduling45,46, the particle swarm optimization algorithm for energy
digital twin framework for production scheduling—which includes optimization47 and the Bayesian network for remaining tool life predic-
a physical flowshop system, an input data module, an optimization tion48. However, research on data-quality improvement, to maximize
module, a digital twin module and an operator interface—was pro- the value of twin data, is relatively scarce, which cannot be neglected
posed to ensure the robustness of the optimal scheduling solutions and should be given more attention.
by considering the uncertainties of a real flowshop22. In addition, a
digital twin framework for the simulation and optimization of pro- Twin interaction
duction systems was proposed to better configure and implement Twin interaction contains the interactions between the twin model and
digital twins in the design and operation stages. In this framework, the physical object, between the twin model and twin data, between
the production twin and machine/process twin were provided with the physical object and twin data, between the physical object and ser-
the details about their components and functions3. vices, between the twin model and services, and between twin data and
Even though conceptual models and reference frameworks are services, thereby forming an organic whole. Serendipitous factors, such
important for providing guidelines for the compositions and the imple- as unexpected stochastic disturbance, frequent insertion order and
mentation process of digital twins, they alone are not enough, as their equipment failures, usually make the virtual space inconsistent with
specific implementation depends on various supporting technologies. practical situations in the industrial scene. This is where twin interac-
tions come into play: the consistency and synchronization between the
Twin perception virtual and physical spaces, which is one of the core features of digital
One of the most important steps of establishing the relationships and twins, can be realized by virtual–real interaction49. Therefore, some
realizing the synchronization between the virtual and physical spaces interaction strategies or methods have been studied in the past, for
is sensing changes of the environment, machines, people and other example, the specification-based state replication method50 and the
elements in the manufacturing systems. This process mainly involves Kalman-filter-as-observer method51. These methods can be mainly
data access, data description, and data transformation from sensors, divided into two steps. First, the differences between virtual and physi-
intelligent terminals and other devices, in which the Industrial Internet cal spaces should be uncovered before the virtual space can be updated
of Things (IoT) becomes a valuable choice23–25. For example, electric accordingly. The rule-based consistency-checking method could be
current sensors, accelerometer sensors and acoustic emission sen- used to achieve this goal by comparing the sensed timed event from the
sors could be used to collect the status data of computer numerical physical manufacturing plant with the digital twins-based estimations52.
control (CNC) machines26. In addition, the data from the CNC system After that, the problem of how to make a virtual space consistent with its
and operation processes—including feed rate, depth of cut, cutter corresponding physical space within an acceptable error and time delay
location of the cutting tools, production data, alarm information and can be studied. For example, the anchor point method has been applied
so forth—are also needed27,28. The collected data are often integrated to realize synchronization of the manufacturing automation system53.
into databases or the cloud for further use through data exchange
protocols, including transmission control protocol/internet protocol29, Applications
open platform communications unified architecture30, MTConnect31 On the basis of the developed models and technologies, some explo-
and automation markup language32. rations and practices have been carried out towards industrial needs.
Current applications have shown digital twins’ priorities in the entire
Twin model construction product and shop-floor life cycle, as shown in Fig. 2.
Twin models are used to depict the key features (such as geometry In the design phase, compared with traditional design methods,
features, physics features and behavior features) of the manufacturing for example, brainstorming, it is much easier to generate new solutions
system and its elements33–35. Different modeling methods—including by using high-fidelity twin models and life-cycle data. The models can
physics-based modeling, data-based modeling and hybrid modeling— support virtual testing with a high level of confidence, which thus can
have been applied in different scenarios36. The accuracy of physics- greatly reduce the costs and time of physical testing54. In addition,
based models relies on the understanding of the physical mechanisms twin models and product life-cycle data could bridge the gap between
and objective laws to a great extent, but this method does not need design and manufacturing to avoid the condition where the design
many data. On the contrary, accurate data-based modeling needs schemes cannot be realized with limited manufacturing capacity.
many parameters and many data with high quality. In reality, some For example, Huang et al. proposed a framework for rapid product
mechanisms are not well enough understood and the available data are development based on digital twins. The digital product prototypes
limited. A hybrid modeling method could take advantage of these two were obtained based on digital trial manufacturing, which contributes
methods, and as such, is regarded as the trend to realize the accurate to improving the manufacturing process during the design stage, and
description of the targets through the fusion of data, knowledge and rotorcraft components were used to validate the performance of the
mechanisms37–39. proposed framework54.

Nature Computational Science | Volume 4 | March 2024 | 169–177 171

Perspective https://doi.org/10.1038/s43588-024-00603-w

Guiding the process of Condition monitoring

Digital twins’ Digital design, testing Digital reconfiguration
implementation and and operation
applications and verification design and verification
configuration optimization

Shop floor Operation and

Design Construction Reconfiguration
life cycle maintenance

Manufacturing
process
Shop floor

Product

Product
Recycling Usage Production Design
life cycle

Providing guidance Operation Process monitoring Digital design, Digital twins’

for disassembly and optimization, and quality management testing and applications
remanufacturing prognostic and verification
health management

Fig. 2 | Digital twin applications in industry from the perspective of product of shop floors and products through integrated simulation, real-time condition
life cycle and shop floor life cycle. Digital twins can be applied in the entire monitoring, prediction, optimization and so forth. In the figure, red is used to
life cycle of both the product and the shop floor, and form a coupling in the represent product-related content; blue is used to represent shop floor-related
manufacturing process. The main role of digital twins is to assist in the design and content; and the green dotted box represents the manufacturing process, in
implementation of products and shop floors and to guarantee the performance which both the products and the corresponding shop floors should be included.

In the production or construction phase, the interaction strategy The effects of digital twins at each of these stages are clear, and
makes it possible to realize the synchronization between the virtual many companies have been working on this space, including GE,
models and practical status, which can support real-time monitoring55. Siemens, IBM and Dassault Systems63. However, there are still two major
Furthermore, in the virtual space, multiple schemes (for example, drawbacks that limit the further development of digital twins. One is
logistics schemes and control schemes) could be simulated to that some existing pitfalls (see below) have led to the improper use of
support decision-making and to improve efficiency56,57. Instructions digital twins, so the full value of digital twins has not been fully realized
could then be given to the elements in the manufacturing systems, yet64. In addition, there are some related technological bottlenecks to
creating a closed loop between the virtual and physical spaces. On overcome. Finally, current applications from academic papers mostly
the basis of this idea, for the uncertain events, information asym- focus on laboratories or pilot production lines rather than practical
metry and abnormal interference in the manufacturing process of factories. More efforts are needed to move digital twins from theory
the shop floor, Fang et al. used digital twins to obtain equipment to practice. Some companies have provided digital twin solutions, but
states, production tasks, interference and other relevant informa- there is still a long way to go for large-scale applications.
tion from the physical shop floor. Shop-floor dynamic scheduling
was realized through the analysis of processing time, worker skill Main pitfalls
level, cost, energy consumption and other data. Thus, the scheduling Overly simplistic or overly complex models
deviation could be effectively reduced and the efficiency and quality Modeling is mainly used to describe the features of physical objects,
of scheduling execution could be improved58. which raises a question about how to determine the right level of detail.
In the operation and maintenance or usage and recycling phases, It goes without saying that the more accurate the models, the better
rich sensing data reflect the current state of machines, production the simulation results. However, overly complex models require a lot of
lines, factories or other objects. By combining historical data, field resources (computing costs and time), which can not only be unneces-
data and twin models, a more accurate prediction can be made59. sary but also not satisfy the timeliness requirement in some situations.
When a machine presents a failure, the fault link, the fault location For example, to minimize the maximum delay time through dynamic
and the fault cause could be determined rapidly by using digital twins. job shop scheduling, the material flow, process flow and information
At the same time, maintenance planning and remote maintenance flow are the focuses of twin models. Besides, the health status, capacity,
could also be realized, preventing serious consequences and reduc- location of equipment and process execution are additional factors
ing overall costs60–62. that should be considered. However, the models of machines’ screws

Nature Computational Science | Volume 4 | March 2024 | 169–177 172

Perspective https://doi.org/10.1038/s43588-024-00603-w

or product structures are not necessary to obtain accurate results and more attention from users and researchers. Through comprehensive
could even be neglected. interactions, the accuracy of models, the value of data and the quality
Under this condition, some people may ask whether simpler of service can all be improved.
models are recommended instead. The answer is, however, ‘no’.
Ideally, we would like to use simple models effectively to solve complex Underused artificial intelligence
problems, but in reality, it is hard to do so. For example, the three- There is a substantial amount of data coming from machines, humans,
dimensional models of a CNC machine tool can reflect only its actions, materials and the environment in the physical space, which brings the
which allows the visualization of the running process. Because the challenge of how to deal with these data during the implementation
physical features (for example, fatigue wear and corrosion fatigue) of digital twins. Notably, artificial intelligence is suitable for mining
are not embodied in three-dimensional models, the prognostics and knowledge from huge amounts of data. Artificial intelligence has been
health management of the machines cannot be assessed. The problems flourishing in the past few years, and while it is still under constant devel-
exposed by overly simplistic and overly complex models or elements opment, it might be mature enough to be applied to some practical
include ambiguous requirements and unbalanced input–output. scenarios. For instance, digital twins can provide dynamic scheduling,
Therefore, it is important to accurately clarify user needs and analyze equipment failure prediction, energy consumption optimization and
specific problems, thus guiding the trade-off of model construction or some other services with high quality by the use of artificial intelligence.
reconstruction between complex models and simple models. In process manufacturing, Priyanka et al. combined digital twins with
machine learning to predict the risk probability rate of an oil pipeline
Not just big data system and to assess its remaining useful life69. In discrete manufactur-
Data are essential to bridge the virtual and physical spaces, which can ing, to reduce the impact caused by uncertainties and unpredictable
drive models to be in silico representations of their physical coun- events, Chen et al. proposed a multi-objective scheduling strategy of
terparts and machines to operate. However, current research pays a fuzzy flexible job shop under the uncertainty of completion time,
more attention to big data and neglects the value of small data to delivery date and equipment maintenance cycle, and the hybrid particle
some extent. The difference between big data and small data lies in swarm optimization algorithm with a variable neighborhood search
the scale, complexity and processing methods of the data. Big data strategy was used to find the optimal scheduling scheme. The effec-
often require the use of distributed computing and artificial intelli- tiveness and feasibility of the proposed method were illustrated in an
gence techniques for processing and analysis, whereas small data can automobile supporting factory70.
be processed and analyzed using traditional statistical analysis and Apart from data mining, artificial intelligence could also be applied
data-mining methods. In some conditions, small data could beat big in other aspects during the digital twin implementation. Considering
data. A small dataset with high quality is more meaningful for specific the limitations of computing power on the shop floor, artificial intel-
industrial services than large observational data of unknown prov- ligence could help solve the problem of optimally allocating comput-
enance. In addition, big data have a relatively poor performance in ing resources. Moreover, artificial intelligence could be embodied
assessments of uncertainty65. Some dangers of model misspecification into edge hardware71, greatly reducing the amount of data transfer
and uncertainty for small data could be obtained, which is helpful for and better meeting the timeliness requirement of digital twins. Some
calibrating the uncertainty66. Neglecting small data may also cause bias recent products based on neural network and large language models,
in big data. Big data usually have a better performance in discovering such as ChatGPT, may provide additional possibilities to digital twins.
rules but have a relatively poor performance in imperceptible changes.
Hence, it is better to integrate the strengths and limitations from both Main challenges
small data and big data approaches into a hybrid approach to support Accurate model construction and comprehensive validation
better industrial services. Moreover, in the process of acquiring big data It is easy to build an accurate model for simple objects. However,
and small data, the strategy of data collection, data transmission, data in many situations, the structure of the processing equipment, the
storage and data analysis should also be seriously evaluated to avoid technological process and the operating environment all may be very
unprofitable or unnecessary data collection. complicated, which brings challenges to modeling. The mechanism
for some processes and coupling relationships may not be clear or
Inadequate interaction accurate enough either, making it difficult to build their corresponding
Interaction is one of the main features of digital twins compared with models. For example, the degree of tool wear is affected by many fac-
traditional offline simulation, and focus should be given to the timeli- tors, such as tool material and shape, working environment and cooling
ness and comprehensiveness of the interaction processes. Timeliness method, cutting parameters (such as cutting speed, feed rate and the
reflects the execution time of the interaction. The higher the real-time material to be machined) and so forth. In addition, the different forms
requirement, the faster the data transmission needs to be. Although of tool wear, including abrasive wear, adhesive wear, diffusion wear,
some new data transmission technologies (for example, 6G67,68) have fatigue wear and chemical wear, also relates to the different factors
been developed, the transmission speed is still limited. Therefore, mentioned above72. It is then hard to depict the process of tool wear
we should not blindly pursue high interaction timeliness, but choose accurately. In addition, many challenges lie in the balance between
different interaction strategies according to each situation. For exam- computation cost and model accuracy: taking all of the aforemen-
ple, some static production elements (such as machine tool bed) do not tioned factors into account would require a substantial number of
need to interact with other elements. However, safety-related elements parameters, bringing huge computing costs. It is worth noticing that
require a high interactive timeliness. Another important considera- the models here are not infinitely accurate but valid on the premise
tion is comprehensiveness, which refers to how much elements and of satisfying customer demand. To tackle this challenge, the fusion
links are considered in the process of interaction. Current applications between physics-based models and data is promising, because the
often focus on the interaction between physical and virtual spaces, or knowledge mined by data can compensate the deficiencies of the
between human and machine. However, we believe that other interac- unknown mechanism to some extent.
tions should also be considered. There are many subelements in virtual Model validation is the main way to evaluate the model perfor-
and physical spaces, and these subelements also need to interact with mance and the degree of confidence, and an indispensable step for
each other, including interactions among virtual elements (such as digital twin applications. One-sided or wrong evaluation results will
models, data and service) and among physical elements (such as the probably mislead the usage of the models and even cause serious
person, machine, material, method and environment), which need consequences. A comprehensive model validation is needed to assist in

Nature Computational Science | Volume 4 | March 2024 | 169–177 173

Perspective https://doi.org/10.1038/s43588-024-00603-w

judging where the models are suitable for. However, there is currently a consensus on the judgement criteria on ‘explainability’, creating an
a lack of related international standards or benchmarks in industry to acceptable benchmark considering subjective and objective factors,
guide the implementation of model validation. and making the trade-off between explainability and performance
should be further developed in the future80.
Data tasks under complex working conditions
Data collection, transmission, storage and processing are the main Security of digital assets
steps to create the value of twin data, and there are challenges in each Twin models and twin data have become important digital assets as
of these steps. business values are often embodied in these models and data. For
example, the twin models of CNC machines contain the information
Data acquisition. Even though there are some common data- of structure, material, hydraulic system, cooling system, lubrication
acquisition methods—including configuration software, data-acqui- system and so forth. Therefore, the security of digital assets is essential.
sition card, sensors and radio-frequency identification devices—there If the security issues are not well addressed, it is difficult to realize the
are still some challenges to be addressed. Some interfaces of machines industrial chain coordination (meaning, cooperation across different
are not open, and some machines do not even have interface to support enterprises). Recently, some promising security technologies have
data collection. Besides, some machines cannot be installed with sen- been developed. Blockchain is one of the potential security solutions
sors, as the installed sensors would affect their performance. In some for disseminating digital assets across the stakeholders through the
complex working conditions—for example, high temperature, high characteristics of decentralization, encryption technology and distrib-
pressure, low temperature, dust and high-radiation flux—there are uted ledger81–83. However, more efforts are still needed for developing
higher requirements for sensors, including security, microminiaturiza- blockchain and addressing other security issues (for instance, digital
tion, high precision, low power consumption and so forth. assets attacked by an outside party).

Data transmission. As more and more data are collected, higher Common industrial software and platform
requirements are put forward for data transmission speed. In addition, Mature industrial software and platforms could greatly enhance digi-
data transmission faces the risk of unauthorized access, data leakage tal twin applications. At present, some companies have developed
and data tampering. Common encryption technologies, including link tools to support digital twin implementation, such as ThingWorx,
encryption, node encryption and end-to-end encryption, can ensure 3DEXPERIENCE, TwinBuilder, AWS IoT TwinMaker, Azure Digital Twins
secure data transmission to some extent, but their performance in data and Simcenter84. However, the software usually cannot meet the needs
protection is still limited73. of researchers, developers and enterprises84. Often, the functions are
rather limited, as companies develop their software based on their
Data storage. The large volume of data makes the data storage burden main foundation and strengths. In addition, file formats between dif-
increasingly unbearable. The software-defined-storage approach could ferent software are generally incompatible. To realize digital twins in
alleviate the above problems, because it utilizes software to manage industry more widely85, a common industrial software or platform is
and control storage resources, contrasting with traditional hardware- needed, which would break the bespoke mode. In this process, how
dependent storage solutions. Software-defined storage abstracts to balance the benefits of different software companies is also a huge
storage functionality from hardware, enabling the management and challenge. Therefore, this vision requires the efforts of the entire digital
configuration of storage resources through software-defined methods. twin community.
Additional storage nodes or devices can also be easily incorporated to
meet growing storage demands74. However, to realize giant techno- Standard of digital twins in industry
logical leaps, the next generation of transformative optical storage An international standard is the foundation for large-scale industrial
technology looks promising75. application of digital twins. Although standardization work has started
(for example, ISO 23247, IPC 2551 and ISO/IEC JTC 1/SC 41), different
Data processing. Both unstructured and structured data are collected organizations have not reached an agreement on the understanding of
in industrial applications. Hence, how to integrate these data formats, digital twins yet, which may result in compatibility and interoperability
ensure data quality and maximize data value becomes a primary task. problems. Therefore, cooperation among standards organizations is
Recently, advanced algorithms have been more widely used in the needed in the future to maintain consistency and avoid redundancies.
field of manufacturing to solve the problem of classification, predic- Moreover, some existing standards about the technologies used in
tion and optimization. However, better algorithms are still needed to digital twins could be still adopted. For example, ISO 29002, IEC 61987
make full use of collected data to meet the needs of different industrial and ISO 13372:2012 could be used in digital twins’ industrial applica-
applications. In addition, the explainability, robustness and fairness of tions to maintain compatibility. Standards on software, development
algorithms should be further improved. To run these algorithms, the processes and implementation processes toward different industrial
time consumption has a close relationship with the computing power, scenarios are still absent.
which would directly affect the timeliness of digital twins. Even though
the technology of multi-core chips could optimize the performance of Potential ethical and privacy concerns
traditional chips, the improvements are in a limited range. The introduction of digital twins inevitably raises privacy and ethical
concerns for users, as digital twins would contain multiple data and
Explainable artificial intelligence for industrial applications models of the whole manufacturing systems, including private data
Many factories, shop floors, production lines and products have a high from individuals. Individuals may be concerned about models and data
requirement for safety and reliability, as they may contain dangerous being abused and misused, thus requesting ownership and control of
articles (for example, crude oil) that are inflammable or explosive. data and models. In addition, some new cybercriminal activities may
While artificial intelligence has been used to improve the effects and also emerge86. Finally, some biases may occur because of incomplete
values of some digital twins’ industrial applications76, it is usually a data or selected algorithms, which would further lead to one-sided con-
‘black box,’ which means that how the models work and the results clusions and unreasonable decisions. To address the aforementioned
are obtained may not be entirely clear77. In addition, it is difficult to issues, related technologies, policies, laws and regulations need to be
assign responsibility once accidents occur. Explainable artificial intel- developed. For example, appropriate information transparency plays
ligence provides a path to address this problem78,79. However, reaching an important role in increasing trust of digital twins. Only when the

Nature Computational Science | Volume 4 | March 2024 | 169–177 174

Perspective https://doi.org/10.1038/s43588-024-00603-w

potential ethical and privacy concerns have been well addressed can 16. Pei, F. Q., Tong, Y. F., Yuan, M. H., Ding, K. & Chen, X. H. The digital
digital twin applications be widely accepted and trusted. twin of the quality monitoring and control in the series solar cell
production line. J. Manuf. Syst. 59, 127–137 (2021).
Conclusion and outlook 17. Magnanini, M. C. & Tolio, T. A. M. A model-based digital twin
Despite the considerable advancements in industrial applications in to support responsive manufacturing systems. CIRP Ann. 70,
recent years, the overall maturity of digital twins remains relatively 353–356 (2021).
low due to insufficient recognition, inaccurate models, incomplete 18. Toothman, M. et al. A digital twin framework for prognostics and
data, inadequate interaction, immature commercial software and a health management. Comput. Ind. 150, 103948 (2023).
fragmentary standard system. In addition, the development of digital 19. Michael, G. & Vickers, J. in Transdisciplinary Perspectives on
twins in industry faces persistent technical challenges that need to be Complex Systems: New Findings and Approaches (eds Kahlen, F. J.
overcome to drive further progress. et al.) 85–113 (Springer, 2017).
Digital twins are not isolated from other advanced technologies 20. Tao, F. et al. Digital twin-driven product design, manufacturing
(such as artificial intelligence), but instead, have a close relationship and service with big data. Int. J. Adv. Manuf. Technol. 94,
with them. In the future, digital twins will potentially combine the 3563–3576 (2018).
theory of reliability, complexity and systems engineering to further 21. Negri, E., Berardi, S., Fumagalli, L. & Macchi, M. MES-integrated
improve their application effect and better solve systemic problems of digital twin frameworks. J. Manuf. Syst. 56, 58–71 (2020).
complex objects. In addition, to make digital twins sustainable, mature 22. Negri, E. et al. Field-synchronized digital twin framework for
business models, such as a technology provider model and an applica- production scheduling with uncertainty. J. Intell. Manuf. 32,
tion service provider model, should be developed. Policy support is also 1207–1228 (2021).
one of the important driving forces for researchers and enterprises at 23. Souza, V., Cruz, R., Silva, W., Lins, S. & Lucena, V. A digital
the same time. In a word, to realize digital twins, collaborative efforts twin architecture based on the Industrial Internet of Things
from all stakeholders involved are essential. technologies. In 2019 IEEE International Conference on Consumer
Electronics (ICCE) 1–2 (IEEE, 2019).
References 24. Gopal, L. et al. Digital twin and IOT technology for secure
1. Tao, F. & Qi, Q. Make more digital twins. Nature 573, 490–491 manufacturing systems. Meas. Sens. 25, 100661 (2023).
(2019). 25. Redelinghuys, A. J. H., Basson, A. H. & Kruger, K. A six-layer
2. Lei, Z. et al. Digital twin based monitoring and control for DC–DC architecture for the digital twin: a manufacturing case study
converters. Nat. Commun. 14, 5604 (2023). implementation. J. Intell. Manuf. 31, 1383–1402 (2020).
3. Ricondo, I., Porto, A. & Ugarte, M. A digital twin framework for the 26. Ghosh, A. K., Ullah, A. S., Teti, R. & Kubo, A. Developing sensor
simulation and optimization of production systems. Procedia CIRP signal-based digital twins for intelligent machine tools. J. Ind. Inf.
104, 762–767 (2021). Integr. 24, 100242 (2021).
4. Hernandez-Boussard, T. et al. Digital twins for predictive oncology 27. Cai, Y., Starly, B., Cohen, P. & Lee, Y.-S. Sensor data and
will be a paradigm shift for precision cancer care. Nat. Med. 27, information fusion to construct digital-twins virtual machine tools
2065–2066 (2021). for cyber-physical manufacturing. Procedia Manuf. 10, 1031–1042
5. Coorey, G., Figtree, G. A., Fletcher, D. F. & Redfern, J. The health (2017).
digital twin: advancing precision cardiovascular medicine. 28. Shahriar, M. R. et al. MTComm based virtualization and
Nat. Rev. Cardiol. 18, 803–804 (2021). integration of physical machine operations with digital-twins
6. Coorey, G. et al. The health digital twin to tackle cardiovascular in cyber-physical manufacturing cloud. In 2018 5th IEEE
disease—a review of an emerging interdisciplinary field. npj Digit. International Conference on Cyber Security and Cloud Computing
Med. 5, 126 (2022). (CSCloud)/2018 4th IEEE International Conference on Edge
7. Shrivastava, C., Berry, T., Cronje, P., Schudel, S. & Defraeye, Computing and Scalable Cloud (EdgeCom) 46–51 (IEEE, 2018).
T. Digital twins enable the quantification of the trade-offs in 29. Wang, K. J., Lee, Y. H. & Angelica, S. Digital twin design for real-
maintaining citrus quality and marketability in the refrigerated time monitoring—a case study of die cutting machine. Int. J. Prod.
supply chain. Nat. Food 3, 413–427 (2022). Res. 59, 6471–6485 (2021).
8. Schrotter, G. & Hürzeler, C. The digital twin of the city of Zurich 30. Luo, W., Hu, T., Zhang, C. & Wei, Y. Digital twin for CNC machine
for urban planning. J. Photogramm. Remote Sens. Geoinf. Sci. 88, tool: modeling and using strategy. J. Ambient Intell. Hum.
99–112 (2020). Comput. 10, 1129–1140 (2019).
9. Aydemir, H., Zengin, U. & Durak, U. The digital twin paradigm for 31. Urbina Coronado, P. D. et al. Part data integration in the shop
aircraft review and outlook. In AIAA SciTech Forum 0553 (AIAA, floor digital twin: mobile and cloud technologies to enable a
2020). manufacturing execution system. J. Manuf. Syst. 48, 25–33 (2018).
10. Coraddu, A. et al. Data-driven ship digital twin for estimating the 32. Bao, J., Guo, D., Li, J. & Zhang, J. The modelling and operations for
speed loss caused by the marine fouling. Ocean Eng. 186, 106063 the digital twin in the context of manufacturing. Enterp. Inf. Syst.
(2019). 13, 534–556 (2019).
11. Mauro, F. & Kana, A. A. Digital twin for ship life-cycle: a critical 33. Koulouris, A., Misailidis, N. & Petrides, D. Applications of process
systematic review. Ocean Eng. 269, 113479 (2023). and digital twin models for production simulation and scheduling
12. Bauer, P., Stevens, B. & Hazeleger, W. A digital twin of Earth for the in the manufacturing of food ingredients and products. Food
green transition. Nat. Clim. Change 11, 80–83 (2021). Bioprod. Process. 126, 317–333 (2021).
13. Uhlemann, T. H.-J., Lehmann, C. & Steinhilper, R. The digital twin: 34. Melesse, T. Y., Pasquale, V. D. & Riemma, S. Digital twin models
realizing the cyber-physical production system for Industry 4.0. in industrial operations: state-of-the-art and future research
Procedia CIRP 61, 335–340 (2017). directions. IET Collab. Intell. Manuf. 3, 37–47 (2021).
14. Wang, K. et al. A review of the technology standards for enabling 35. Tao, F., Xiao, B., Qi, Q., Cheng, J. & Ji, P. Digital twin modeling.
digital twin. Digit. Twin 2, 4 (2022). J. Manuf. Syst. 64, 372–389 (2022).
15. Lo, C. K., Chen, C. H. & Zhong, R. Y. A review of digital twin in 36. Rasheed, A., San, O. & Kvamsdal, T. Digital twin: values,
product design and development. Adv. Eng. Inform. 48, 101297 challenges and enablers from a modeling perspective. IEEE
(2021). Access 8, 21980–22012 (2020).

Nature Computational Science | Volume 4 | March 2024 | 169–177 175

Perspective https://doi.org/10.1038/s43588-024-00603-w

37. Hürkamp, A. et al. Combining simulation and machine learning as 57. Aheleroff, S., Xu, X., Zhong, R. Y. & Lu, Y. Digital twin as a service
digital twin for the manufacturing of overmolded thermoplastic (DTaaS) in Industry 4.0: an architecture reference model. Adv.
composites. J. Manuf. Mater. Process 4, 92 (2020). Eng. Inform. 47, 101225 (2021).
38. Tripura, T., Desai, A. S., Adhikari, S. & Chakraborty, S. Probabilistic 58. Fang, Y. et al. Digital-twin-based job shop scheduling
machine learning based predictive and interpretable digital twin toward smart manufacturing. IEEE Trans. Ind. Inform. 15,
for dynamical systems. Comput. Struct. 281, 107008 (2023). 6425–6435 (2019).
39. Balu, A., Sarkar, S., Ganapathysubramanian, B. & Krishnamurthy, A. 59. Liu, D., Du, Y., Chai, W., Lu, C. & Cong, M. Digital twin and data-
Physics-aware machine learning surrogates for real-time driven quality prediction of complex die-casting manufacturing.
manufacturing digital twin. Manuf. Lett. 34, 71–74 (2022). IEEE Trans. Ind. Inform. 18, 8119–8128 (2022).
40. Tabar, R. S., Wärmefjord, K., Söderberg, R. & Lindkvist, L. Efficient 60. Aivaliotis, P., Georgoulias, K. & Chryssolouris, G. The use of digital
spot welding sequence optimization in a geometry assurance twin for predictive maintenance in manufacturing. Int. J. Comput.
digital twin. J. Mech. Des. 142, 102001 (2020). Integr. Manuf. 32, 1067–1080 (2019).
41. Namiot, D., Pokusaev, O., Kupriyanovsky, V. & Zhabitskii, M. Digital 61. Errandonea, I., Beltrán, S. & Arrizabalaga, S. Digital twin
twins and discrete-event simulation systems. Int. J. Open Inf. for maintenance: a literature review. Comput. Ind. 123,
Technol. 9, 70–75 (2021). 103316 (2020).
42. Morabito, L., Ippolito, M., Pastore, E., Alfieri, A. & Montagna, F. 62. van Dinter, R., Tekinerdogan, B. & Catal, C. Predictive
A discrete event simulation based approach for digital twin maintenance using digital twins: a systematic literature review.
implementation. IFAC Pap. 54, 414–419 (2021). Inf. Softw. Technol. 151, 107008 (2022).
43. Ganguli, R. & Adhikari, S. The digital twin of discrete dynamic 63. Ribeiro da Silva, E., Assad Neto, A. & Nielsen, C. P. in The Future
systems: initial approaches and future challenges. Appl. Math. of Smart Production for SMEs: A Methodological and Practical
Model. 77, 1110–1128 (2020). Approach Towards Digitalization in SMEs (eds Madsen, O. et al.)
44. Söderberg, R., Wärmefjord, K., Carlson, J. S. & Lindkvist, L. Toward 343–348 (Springer, 2023).
a digital twin for real-time geometry assurance in individualized 64. Wanasinghe, T. R. et al. Digital twin for the oil and gas industry:
production. CIRP Ann. 66, 137–140 (2017). overview, research trends, opportunities, and challenges. IEEE
45. Yan, Q., Wang, H. & Wu, F. Digital twin-enabled dynamic Access 8, 104175–104197 (2020).
scheduling with preventive maintenance using a double-layer 65. Faraway, J. J. & Augustin, N. H. When small data beats big data.
Q-learning algorithm. Comput. Oper. Res. 144, 105823 (2022). Stat. Probab. Lett. 136, 142–145 (2018).
46. Guo, X., Peng, G. & Meng, Y. A modified Q-learning algorithm for 66. Yu, J., Song, Y., Tang, D. & Dai, J. A digital twin approach based
t path planning in a digital twin assembly system. Int. J. Adv. on nonparametric Bayesian network for complex system health
Manuf. Technol. 119, 3951–3961 (2022). monitoring. J. Manuf. Syst. 58, 293–304 (2021).
47. Chen, R., Shen, H. & Lai, Y. A metaheuristic optimization algorithm 67. Dang, S. et al. What should 6G be? Nat. Electron. 3, 20–29
for energy efficiency in digital twins. Internet Things Cyber Phys. (2020).
Syst. 2, 159–169 (2022). 68. Saad, W., Bennis, M. & Chen, M. A vision of 6G wireless systems:
48. Bazaz, S. M., Lohtander, M. & Varis, J. The prediction method of applications, trends, technologies, and open research problems.
tool life on small lot turning process—development of digital twin IEEE Netw. 34, 134–142 (2020).
for production. Procedia Manuf. 51, 288–295 (2020). 69. Priyanka, E. B., Thangavel, S., Gao, X.-Z. & Sivakumar, N. S. Digital
49. Zhang, H., Qi, Q., Ji, W. & Tao, F. An update method for digital twin for oil pipeline risk estimation using prognostic and machine
twin multi-dimension models. Robot Comput. Integr. Manuf. 80, learning techniques. J. Ind. Inf. Integr. 26, 100272 (2022).
102481 (2023). 70. Chen, Z., Zou, J. & Wang, W. Digital twin-oriented collaborative
50. Eckhart, M. & Ekelhart, A. A specification-based state replication optimization of fuzzy flexible job shop scheduling under multiple
approach for digital twins. In Proc. 2018 Workshop on Cyber- uncertainties. Sādhanā 48, 78 (2023).
Physical Systems Security and Privacy 36–47 (Association for 71. Singh, R. & Gill, S. S. Edge AI: a survey. Internet Things Cyber Phys.
Computing Machinery, 2018). Syst. 3, 71–92 (2023).
51. Akbarian, F., Fitzgerald, E. & Kihl, M. Synchronization in digital 72. Olortegui-Yume, J. A. & Kwon, P. Y. Tool wear mechanisms in
twins for industrial control systems. Preprint at https://arxiv.org/ machining. Int. J. Mach. Mach. Mater. 2, 316–334 (2007).
abs/2006.03447 (2020). 73. Suo, S. et al. Encryption technology in information system
52. Seok, M. G., Tan, W. J., Cai, W. & Park, D. Digital-twin consistency security. In 3rd International Conference on Mechatronics
checking based on observed timed events with unobservable Engineering and Information Technology (ICMEIT 2019) 495–499
transitions in smart manufacturing. IEEE Trans. Ind. Inform. 19, (Springer, 2019).
6208–6219 (2023). 74. Darabseh, A. et al. SDStorage: a software defined storage
53. Talkhestani, B. A., Jazdi, N., Schloegl, W. & Weyrich, M. experimental framework. In 2015 IEEE International Conference on
Consistency check to synchronize the digital twin of Cloud Engineering 341–346 (IEEE, 2015).
manufacturing automation based on anchor points. Procedia 75. Gu, M., Li, X. & Cao, Y. Optical storage arrays: a perspective for
CIRP 72, 159–164 (2018). future big data storage. Light. Sci. Appl. 3, e177 (2014).
54. Huang, S., Wang, G., Lei, D. & Yan, Y. Toward digital validation for 76. Lv, Z. & Xie, S. Artificial intelligence in the digital twins: state of the
rapid product development based on digital twin: a framework. art, challenges, and future research topics. Digit. Twin 1, 12 (2022).
Int. J. Adv. Manuf. Technol. 119, 2509–2523 (2022). 77. Carabantes, M. Black-box artificial intelligence: an
55. Qamsane, Y. et al. A unified digital twin framework for real-time epistemological and critical analysis. AI Soc. 35, 309–317 (2020).
monitoring and evaluation of smart manufacturing systems. In 78. Blazek, P. J. & Lin, M. M. Explainable neural networks that simulate
2019 IEEE 15th International Conference on Automation Science reasoning. Nat. Comput. Sci. 1, 607–618 (2021).
and Engineering (CASE) 1394–1401 (IEEE, 2019). 79. Lundberg, S. M. et al. From local explanations to global
56. Agostino, Í. R. S., Broda, E., Frazzon, E. M. & Freitag, M. in understanding with explainable AI for trees. Nat. Mach. Intell. 2,
Scheduling in Industry 4.0 and Cloud Manufacturing 56–67 (2020).
(eds Sokolov, B. et al.) 39–60 (Springer, 2020); https://doi.org/ 80. Gunning, D. et al. XAI—explainable artificial intelligence. Sci.
10.1007/978-3-030-43177-8_3 Robot. 4, eaay7120 (2019).

Nature Computational Science | Volume 4 | March 2024 | 169–177 176

Perspective https://doi.org/10.1038/s43588-024-00603-w

81. Khan, A., Shahid, F., Maple, C., Ahmad, A. & Jeon, G. Toward smart Competing interests
manufacturing using spiral digital twin framework and twinchain. The authors declare no competing interests.
IEEE Trans. Ind. Inform. 18, 1359–1366 (2022).
82. Suhail, S. et al. Blockchain-based digital twins: research Additional information
trends, issues, and future challenges. ACM Comput. Surv. 54, Correspondence should be addressed to Fei Tao.
240:1–240:34 (2022).
83. Yaqoob, I. et al. Blockchain for digital twins: recent advances and Peer review information Nature Computational Science
future research challenges. IEEE Netw. 34, 290–298 (2020). thanks Michael Grieves, George Q. Huang and Jay Lee for their
84. Tao, F. et al. makeTwin: a reference architecture for digital twin contribution to the peer review of this work. Primary Handling
software platform. Chin. J. Aeronaut. https://doi.org/10.1016/j. Editor: Fernando Chirigati, in collaboration with the Nature
cja.2023.05.002 (2023). Computational Science team.
85. Niederer, S. A. et al. Scaling digital twins from the artisanal to the
industrial. Nat. Comput. Sci. 1, 313–320 (2021). Reprints and permissions information is available at
86. Suhail, S., Jurdak, R. & Hussain, R. Security attacks and www.nature.com/reprints.
solutions for digital twins. Preprint at https://doi.org/10.48550/
arXiv.2202.12501 (2023). Publisher’s note Springer Nature remains neutral with regard
to jurisdictional claims in published maps and institutional
Acknowledgements affiliations.
This work is financially supported by the National Natural Science
Foundation of China (NSFC) under grants 52120105008 and 52275471, Springer Nature or its licensor (e.g. a society or other partner) holds
the National Key Research and Development Program of China exclusive rights to this article under a publishing agreement with
under grant 2020YFB1708400, and the New Cornerstone Science the author(s) or other rightsholder(s); author self-archiving of the
Foundation through the XPLORER PRIZE. accepted manuscript version of this article is solely governed by the
terms of such publishing agreement and applicable law.
Author contributions
All authors contributed to the writing and editing of this paper. © Springer Nature America, Inc. 2024

Nature Computational Science | Volume 4 | March 2024 | 169–177 177