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

Introduction

• Digital twins are virtual product models intended to combine twinning, simulation, real-time monitoring, and analytics, enabling virtual prototype testing without disrupting physical operations.The paper describes Digital Twin Prototypes and Digital Twin Instances as two types of digital twins.Source: §1 · pp. 1–2
• Adoption is impeded by dependence on evolving IoT, big data, and machine learning technologies, domain expertise, and the lack of a universally accepted definition or implementation standards.Monitoring and data collection rely on IoT and enterprise information systems, while analytics relies on big data and machine learning; implementations must also be tailored to each domain.Source: §1 · p. 2
• The paper examines technology dependence, domain effects, the gap between ideal and practical digital twins, current limitations, prior models, and debates over definition and implementation.It calls for a comprehensive theoretical specification through a reference framework that can lead to an implementation and evaluation methodology.Source: §1 · p. 2
• Because documented implementation methods are lacking, the paper assumes good inter-connectivity and compatibility among IoT, data, simulation, real-time synchronisation, and machine learning.The authors state that the lack of methodology makes it infeasible to fully assess how these components affect implementation.Source: §1 · p. 2

2. Previous Work

• Reviews often offered domain-specific insights: categorized production-science work as Digital Twin, Shadow, or Model, while examined big data and data models in manufacturing.General review compared 87 applications and proposed a classification scheme, but excluded white papers and industrial works.Source: §2 · pp. 2–3
• Practical implementations received little attention because domain dependence and reliance on multiple technologies make a sufficiently detailed yet universal DT model difficult to formulate.Tao, Zhang, Liu and Nee were among the few authors to explore specific practical implementations.Source: §2 · p. 3
• Earlier reviews raised questions about a cross-domain DT definition, case studies and reference models, and DT’s contributions to IoT and IS.They also considered whether DT covers the product lifecycle or product alone, and the roles of big data, data models, and data-related issues.Source: §2 · p. 3

2.1. How this review differs

• Rather than categorize literature, this review examines practical DT implementations to identify root causes of implementation challenges, limitations, and new properties.It aims to advance DT theory and practical applications.Source: §2.1 · p. 3, §2 · p. 3
• Its novel questions address a universal reference framework, the concept–implementation gap, real-time machine learning and data, adoption barriers, domain effects, and implementation success.The review also examines how domains shape DT implementation.Source: §2.1 · p. 3

3. Digital Twin

• The concept dates to the early 2000s, but industrial implementation followed digital data storage, embedded sensors, and realistic virtual testing.The paper describes early aerospace use and later uptake in manufacturing and by companies including IBM, Siemens, and GE.Source: §3 · pp. 3–4

3.2. Components of DT

• The authors integrate earlier partial accounts into a holistic definition of elementary and imperative DT components, aiming to generalize them across domains.They note that requirements evolve with dependent technologies such as machine learning, big data, and cybersecurity.Source: §3.2 · p. 4, Table 1 · p. 5

3.2.1. Elementary Components

• A DT requires a physical asset, a digital asset, and a two-way synchronized relation for information flow between them.The physical asset may be a product or a product lifecycle.Source: §3.2.1 · p. 4

3.2.2. Imperative Components

• Imperative components include IoT, continuous data, machine learning, security, and performance evaluation; removing any component voids DT functionality and uniqueness.These components require high-fidelity information flow, big-data tools, joint optimization, and evaluation metrics and methods.Source: §3.2.2 · pp. 4–6

3.3. Properties of a DT

• The paper presents DT properties as what makes a digital twin of an existing physical entity more than a simple “digital” “twin.”Source: §3.1 · p. 4, §3.3 · p. 6

3.3.1. Necessary Properties

• Necessary DT properties are real-time connection, self-evolution, continuous machine-learning analysis, time-series data, and domain-specific services.Self-evolution enables real-time learning and adaptation with feedback; synchronization frequency can depend on the application’s update scenario.Source: §3.3.1 · p. 6

3.3.2. Dynamic Properties

• DTs can vary in autonomy and synchronize continuously or at intervals; the framework allows a twin of either a physical asset or its production lifecycle.The paper says a DT differs from related technologies through the presence of all its components.Source: §3.3.2 · p. 6, §3.4 · p. 7

3.4. How is DT di

• Table 3 compares DT with existing technologies.Source: Table 3 · p. 7

3.5. A brief overview of other similar concepts that preceded DT

• Digital Models exchange data manually without showing real-time state, while Digital Shadows copy physical state one-way; DT integrates data flow to reflect the physical object.Source: §3.5 · p. 7
• Semantic Virtual Factory Data Models represent factory entities in manufacturing and industry, but unlike DT they lack real-time synchronization.Source: §3.5 · p. 7
• Product Avatar distributes and decentralizes product information management, but lacks feedback and may represent only parts of a product.Source: §3.5 · p. 8
• Digital Product Memory captures information about a specific physical part and is described as an instantiation of DT, which extends semantic or digital product memory.Source: §3.5 · p. 8
• DT extends Intelligent Product by using IoT, Big Data, and Machine Learning technologies that were lacking in that earlier concept.Source: §3.5 · p. 8
• Holons, an early computer-integrated manufacturing tool, formed the basis for the technologies described above.Source: §3.5 · p. 8
• The authors conclude that DT offers a collection of novel and exclusive applications.Source: §3.5 · p. 8

4.1. Theoretical models

• Theoretical models range from domain-specific PLM and manufacturing frameworks to generalized DT dimensions, while definitions differ on whether the twin represents a product or its lifecycle.Tao et al. also discuss DT use in product design and manufacturing; the reviewed literature does not fully bridge theory and practice.Source: §4.1 · pp. 8–9
• Autonomous-systems work frames DTs as real-time data and knowledge sources for simulating decisions, but its examples use them mainly as memory buffers rather than for analytics or machine learning.The work also positions DT alongside modularity, connectivity, and autonomy as drivers of future manufacturing.Source: §4.1 · pp. 8–9

4.2. Practical models and their domain-dependence

• Practical DT descriptions are scant and often vague for complex systems; sector needs differ, and implementation requires domain expertise, including human knowledge for adaptive intelligence.CPS, aerospace, manufacturing, and robotics have different expected uses, from failure avoidance and maintenance to simulation and control optimization.Source: §4.2 · pp. 9–10
• Aerospace DTs can model extreme conditions, forecast health and mission success, and support in-flight recommendations, but implementation is difficult amid many parameters; aircraft identity and software interoperability add challenges.Ríos et al. note that commercial aircraft may have more than half a million component references, complicating a unique physical-to-digital mapping.Source: §4.2 · p. 9
• Machine-learning use in DT is described as nearly nonexistent; one petrochemical proof of concept details data preprocessing and system integration, but evaluates its prediction model rather than the DT itself.Prediction criteria include MAR, RMSE, VIR, and PCC.Source: §4.2 · p. 10

4.2.1. Similarity

• Across domains, DT components are expected to remain the same, but implementation feasibility, fidelity, evaluation metrics, software compatibility, expertise needs, and adoption barriers vary.The section also raises compatibility of domain-specific IoT, data, and machine-learning components, and whether regulation and standards impede adoption.Source: §4.2.1 · pp. 10–11

4.3. Industrial Implementations

• The section notes that companies are investing in DT technology, providing DT software, or using DT functionality.Source: §4.3 · p. 11

• Signify Philips is exploring digital twins for lighting by digitising lighting, claiming benefits for emergency services, real-time monitoring, and predictive maintenance.Source: §1 · p. 11

2. Providing DT as a service

• Philips offers healthcare digital twins that warn of technical issues in MRI and CT systems, potentially reducing downtime in clinical spaces.The systems are intended to provide early warning of technical issues.Source: §2 · p. 11
• IBM applies digital twins to monitor and improve Port of Rotterdam efficiency and offers PLM software, while Siemens has power-grid and Formula 1 implementations,,.Source: §2 · p. 11
• Digital-twin implementation software is offered by providers including Dassault Systemes, AnyLogic, Ansys, Visualiz, PwC, Bosch, SAP, and Azure, though their use of machine learning is unclear.Dassault Systemes’ 3DEXPERIENCE has also been used to build a digital twin of Singapore,.Source: §2 · p. 11
• Oracle offers a cloud-based digital-twin simulator, GE offers Predix and holds two digital-twin patents,, and Eclipse Ditto implements IoT-based models.Source: §2 · p. 11

3. Using DT for own use

• DHL’s Tetra Pak warehouse DT in Singapore provides real-time monitoring, while BP’s APEX creates virtual copies of production systems; neither example is reported as global in scope.The authors note that machine-learning use in both systems is unclear and that more concrete analysis of implementation scale is needed.Source: §3 · p. 11

4.4. The gap between the ideal DT and practical DT

• A practical DT may fall short of the ideal because its required IoT, big-data, or machine-learning technologies may not yet be available, while cost and resource constraints can widen the gap.Further assessment depends on access to actual DT software and its initial design model.Source: §4.4 · p. 11
• Implementations’ success is difficult to assess without implementation details and evaluation metrics, and a unified standard DT architecture is needed to distinguish full DTs from partial implementations.The paper notes that claimed uses include data handling, real-time monitoring, simulation testing, and optimisation, but systems lacking components defined in Section 3.2 do not fully meet its conceptual definition.Source: §4.4 · p. 11

4.5. Machine Learning in DT

• Real-time machine learning distinguishes a DT from a simulator or monitoring tool by enabling prediction of asset behavior in unforeseen conditions and prediction of impending problems.Machine learning also enables self-evolving DTs that improve from prediction outcomes and can support resilience to data or model errors; Cronrath et al. use reinforcement learning to resolve such errors.Source: §4.5 · p. 12
• Published DT machine-learning work is scant: a petrochemical proof of concept uses several algorithms [53] but omits its implementation methodology and real-time feedback details, while another work calls simulation without real-time connection a DT approach.The paper says the extent of machine-learning use by companies such as IBM, SAP, and Siemens is uncertain, and identifies synchronization and joint optimization as technical challenges.Source: §4.5 · p. 12

4.6. Big Data in DT

• DT data can be high-dimensional, fragmented by differing sampling frequencies, and difficult to collect and synchronize across connected and unconnected components.The paper calls for practical big-data frameworks and time-series processing methods; one proposed method generates same-frequency series, while other work addresses preprocessing, mining, optimisation, PLM data inconsistencies, and missing data using domain knowledge.Source: §4.6 · pp. 12–13
• Across domains, DT data challenges include gathering IoT data, resourcing storage and preprocessing, reducing high-dimensional data, and standardizing devices and connections for synchronization.Blockchain may help data integrity in health, finance, and cross-industry settings, but increases data-handling complexity and adds an integration and optimization challenge.Source: §4.6 · p. 13

5. Industries where DT can be majorly beneficial

• DTs can replace costly, resource-intensive, time-consuming physical prototypes and simulate extreme tests that are difficult or impossible to perform in labs.Examples include aerospace, supply chain, manufacturing, and PHM.Source: §5 · p. 13
• DTs can monitor emergent behavior in real time, predict imminent problems, and support rapid mitigation; they can also enable holistic analytics across many interdependent sub-components.These capabilities are especially useful in health systems, supply chains, and large organizations with multiple sub-domains where information may remain fragmented.Source: §5 · p. 13
• Despite requiring expertise and resources across multiple technologies, DTs can yield cost reductions from a one-time investment, shorten design cycles, and save prototyping costs, resources, and time.The paper suggests DT-enabled cost reductions could serve as a performance metric for profit-oriented companies; DTs can also predict dangers in time for mitigation.Source: §5 · p. 13

6.1. Challenges

• High-dimensional IoT data collection and preprocessing, together with high-fidelity two-way synchronization, are resource-intensive challenges for large-scale digital twins.Time-series data from numerous devices must be collated by frequency and prepared for machine learning; synchronization requires resources and high-stream IoT connectivity.Source: §6.1 · p. 14
• Simulation software requires programmers, developers, and domain experts, while joint optimization must combine subproblems to account for all physical-asset components.Simulation-based optimization can provide faster, more efficient solutions, but conventional analytical solutions are often entered into simulation software manually rather than optimized there using mechanisms such as reinforcement learning.Source: §6.1 · p. 14
• Digital-twin implementations remain mainly at city or country scale, while interoperability with production software can delay implementation.Larger-scale implementations are desirable for systems such as supply-chain networks and logistics; industries also rely on multiple inventory and product-management software systems.Source: §6.1 · p. 14
• Deep-learning integration requires computational resources, expertise, and research; continual and federated learning appear promising but require further research [85].Managing high-dimensional data alongside industry software and the necessary deep-learning expertise and equipment is tedious.Source: §6.1 · p. 14
• Cross-partner and IoT security concerns are inevitable, and leakage of real-time monitoring data can be hazardous to a firm.Digital twins may operate across multiple industrial partners and inventory sites.Source: §6.1 · p. 14

6.2. Limitations

• Digital twins can be costly for short-lived projects, time-consuming to assemble, and require continued research investment as dependent technologies evolve.Complexity arises from interoperability, real-time tools, joint optimization, and big-data resources; long-term use may add costs through updates to IoT, big data, and machine learning.Source: §6.2 · p. 14

Open questions

• DT performance needs quantitative, domain-dependent and domain-independent metrics, including uncertainty measures, to assess suitability and enable self-evolving systems.Aircraft-specific error metrics were proposed in; quantifying error also helps managers judge confidence in DT outputs and whether to rely on them.Source: §7 · p. 15
• Noisy data and incorrect information models can cause harmful predictions or test results, so resilient DTs need error correction and traceability.Machine learning could support anomaly detection of misfed data, while output-generation logs can help check results and improve models; humans remain important in deciding whether to follow DT recommendations.Source: §7 · p. 15

Future steps

• A formal definition of DT is needed to clarify the concept and build consensus around a universally accepted definition.Source: §7 · p. 15
• Identifying suitable IoT standards for data capture, sharing, synchronization, and monitoring could improve DT acceptance and adoption; Lu et al. also recognize a need for a data communication standard.Source: §7 · p. 15
• Enterprise- and global-level regulations are needed to govern DT data use across collaborating companies and sites subject to different national laws.Source: §7 · p. 15
• Liaising with domain experts can ease DT design, implementation, and acceptance in new sectors by incorporating domain-specific requirements.Programmers and developers can manage implementation once a design is ready.Source: §7 · p. 15

• The proposed DT reference model defines essential components, including a bijective relation with the physical asset and testing and security components for evaluating system performance and integrity.It is based on prior work and is intended to capture the essence of a digital twin.Source: §8 · p. 15
• Regulations and security mechanisms are imperative for widespread DT adoption because its components may span collaborators and industry partners; federated learning is a promising approach to data-sharing concerns.Source: §8 · p. 15
• DT combines simulation, autonomy, agent-based modeling, machine learning, prototyping, optimisation, and big data, with subsystems tailored and prioritised to application-domain needs.Source: §8 · pp. 15–16
• Technical and domain-dependent challenges remain, and advances in DT’s component technologies can hinder development, despite the potential for seamless IoT, machine-learning, and data integration to produce an efficient DT.Source: §8 · p. 16