The emerging role of digital twin technologies in computer engineering research: a systematic literature review

Introduction

Digital twin technology has rapidly moved from a buzzword to a core concept in the design of cyber-physical systems, intelligent networks, and data-driven automation. In general terms, a digital twin is a computational model that maintains a dynamic, data informed representation of a physical asset, process, or system, enabling monitoring, analysis, prediction, and control in near real-time (Grieves and Vickers, 2017; Tao et al., 2019; Fuller et al., 2020). Digital twins are now being explored in domains such as manufacturing, energy, transportation, smart cities, and healthcare, and are increasingly viewed as a key building block for large scale cyber-physical infrastructures (Qi and Tao, 2018; Jones et al., 2020; Yang et al., 2021). However, many existing reviews and application studies emphasize mechanical, industrial, and power systems perspectives, while the computer engineering foundations that make these technologies possible receive much less explicit and systematic attention (Tao et al., 2019; Mashaly, 2021).

Computer engineering plays a central role in realizing digital twins in practice, since twins are instantiated on embedded platforms, networks, and software stacks that must satisfy tight constraints on performance, reliability, and security. Issues such as hardware-software co-design, real-time scheduling, edge and cloud integration, networking for future 5G and 6G systems, on device artificial intelligence, and cybersecurity are not just implementation details; they shape what kinds of digital twin capabilities are feasible and how they behave in real deployments (Fuller et al., 2020; Yang et al., 2021). Yet the literature that addresses these topics is distributed across multiple venues and subfields, often with different terminology and problem framings, which makes it difficult to see how existing work collectively defines the state of digital twin technology from a computer engineering perspective.

Positioning relative to existing surveys

A substantial body of survey and review literature already exists on digital twin technology. To clarify how the present review differs from and complement these prior works, we examined four representative and highly cited surveys, each of which covers a distinct facet of the digital twin landscape.

• Mihai et al. (2022) provide a broad survey of digital twin enabling technologies, challenges, trends, and future prospects in their IEEE Communications Surveys and Tutorials paper. Their review organizes the field around communication, computation, and storage technologies but does not systematically analyze how computer engineering design constraints, such as real-time scheduling, hardware-software co-design, or embedded platform selection, influence what is actually achievable in digital twin deployments.
• Alcaraz and Lopez (2022) offer a comprehensive survey of security threats facing digital twins, also published in IEEE Communications Surveys and Tutorials. While their threat-oriented taxonomy is valuable, the review is organized by attack surface and security mechanism rather than by the computer engineering subsystems (networking stacks, edge nodes, embedded controllers) that must implement those defenses in practice.
• Xu et al. (2023) survey digital twins for the industrial Internet of Things, covering applications, technologies, and tools. Their scope centers on industrial IoT deployments and tool ecosystems, and they provide a useful technology stack decomposition. However, their analysis does not extend into how computer engineering teams navigate trade-offs in platform design, real-time control, or resource-constrained AI, nor does it address digital twin testing and validation as a computer engineering concern.
• Lim et al. (2020) provides an early and influential state-of-the-art survey examining digital twin techniques from engineering product lifecycle management and business innovation perspectives. Their review is anchored in manufacturing and product lifecycle contexts and does not address digital twin architectures for networking, edge computing, or cyber-physical systems more broadly.

Table 1 summarizes the scope and focus of these four surveys alongside the present review, highlighting where each makes its primary contributions and where gaps remain. As Table 1 shows, existing surveys are strong on technologies, threats, industrial tools, and business models, but none of them systematically examine the computer engineering substrate that determines whether a digital twin can actually run reliably, securely, and at scale on real hardware and networks. This review addresses that gap.

TABLE 1. Comparison of the present review with existing digital twin surveys.

Research question and contributions

To address this gap, we conducted a systematic literature review on digital twin technology in computer engineering. Using a structured search and screening process guided by the PRISMA 2020 framework (Page et al., 2021), we identified 45 peer-reviewed publications that explicitly engage with digital twins in computer engineering contexts. We synthesized these studies using inductive thematic analysis and organized them into six themes: computer engineering foundations of digital twin systems, networked and distributed infrastructure for digital twins, computer engineering for intelligent digital twins, security and dependability in digital twin-enabled cyber-physical systems, modeling and testing of digital twin systems, and application domains and use cases.

This review makes three specific contributions that distinguish it from prior work:

• A computer engineering systems perspective on digital twins. Rather than cataloging digital twin technologies by domain (manufacturing, energy, healthcare) or by threat model, this review organizes findings around the computer engineering subsystems that must work together for a digital twin to function: embedded platforms, networking and distributed computing, AI/ML pipelines, security architectures, and testing/validation infrastructure. This perspective reveals cross-cutting design tensions (e.g., fidelity vs. real-time budget, adaptivity vs. verification complexity) that domain-specific surveys do not surface.
• Identification of under-researched computer engineering challenges. The thematic synthesis reveals that while conceptual frameworks and prototype demonstrations are abundant, few studies rigorously evaluate digital twin implementations on constrained computing platforms, address interoperability as a measurable engineering target, or provide reusable validation artifacts (testbeds, benchmarks, reference implementations). These gaps are specific to computer engineering practice and are not highlighted in existing surveys.
• A research agenda grounded in CE design trade-offs. The discussion and implications sections translate thematic findings into concrete research directions, including hardware-aware digital twin design, safe continual-learning pipelines, standardized DT testing methodologies, and cross-domain interoperability frameworks, all framed in terms of the engineering constraints that computer engineering researchers and practitioners face.

This review aims to answer the overall research question: What is the current state of research on digital twin technology in computer engineering? In doing so, it summarizes the main topics studied, highlights trends and gaps in the research, and offers recommendations to guide future work on digital twins within computer engineering.

Methods

We used the PRISMA 2020 framework to structure the review process, as it provides a transparent and replicable approach for identifying, screening, and including studies in a systematic review (Page et al., 2021). Consistent with guidance for systematic reviews in engineering education and related fields (Borrego et al., 2014), our process involved three phases: identification, screening, and inclusion.

Search strategy

We searched seven electronic databases: Google Scholar, Web of Science, IEEE Xplore, Compendex/Engineering Village, ERIC via EBSCOhost, ScienceDirect, and Wiley Online Library. We used two search strings: “digital twin” AND “computer engineering” and “digital twin” AND (“IoT” OR “edge computing” OR “cyber-physical system*”). Searches were limited to journal articles and full-length conference papers. Results were exported, combined, and deduplicated in a single spreadsheet.

Screening and inclusion criteria

Articles were excluded if they met any of the following conditions: (EC1) published before 2015; (EC2) not related to computer or electrical engineering; (EC3) not truly related to digital twin technology that couples a digital and physical representation in a cyber-physical systems context; (EC4) classified as work-in-progress, short papers, posters, or extended abstracts; (EC5) not written in English; (EC6) not available as open access; or (EC7) not clearly situated within the electrical and computer engineering domain. EC6 was applied to ensure that every included article could be independently retrieved, fully inspected, and transparently coded during title, abstract, and full-text screening. The open-access requirement was not used as a proxy for study quality; rather, it supported replicability of the review process by allowing verification of methods, evidence, figures, tables, and claims across the final sample. However, we recognize that this decision may have excluded influential paywalled studies from major venues and publishers, including IEEE, ACM, and Springer, and may therefore have introduced selection bias toward openly available literature. This limitation is revisited in the Limitations section. Articles satisfying none of the exclusion criteria were retained. The complete selection process is summarized in Figure 1.

The database searches returned 541 records. After deduplication and multi-stage screening of titles, abstracts, and full texts against EC1-EC7, 45 articles remained in the final sample. Articles that used “digital twin” only as a synonym for simulation, as well as CPS studies without a digital twin construct, were excluded during full-text review.

Analysis approach

We conducted an inductive thematic analysis following Braun and Clarke’s (2006) six-phase approach. For each article, we recorded: year of publication, venue, country of first author, engineering domain, application area, type of contribution, and how the digital and physical components were coupled. After generating initial codes capturing recurring computer engineering concerns (embedded/real-time systems, networking infrastructure, AI-enabled behavior, security, modeling/simulation, domain applications), we iteratively refined these codes into six themes that organize the findings. The codes, code description, and the exemplars are shown in Table 2. Descriptive statistics summarize publication trends, followed by a qualitative synthesis across the six themes.

TABLE 2. Codes, code description, and exemplars.

Findings

Descriptive overview of publication trends

The 45 articles span 2017 to early 2026, with 39 (87%) published between 2022 and 2025, reflecting the recent emergence of digital twins as a distinct research topic within computer engineering. Publication counts rose from one article in 2017 to a peak of 13 in 2025, with 2 early-view articles in 2026 (Figure 2). Nearly all articles (44 of 45, 98%) appeared in peer-reviewed journals distributed across computer engineering, networking, and CPS outlets, with Future Internet (3 articles) as the most common venue. First authors represented 18 countries, led by the United States and China (6 each, 13.3%), India (5, 11.1%), and the United Kingdom (4, 8.9%).

Among the 20 articles with clearly specified application domains, power and energy systems dominated (55%), followed by Industry 4.0/manufacturing/maintenance (20%), networking/XR/smart retail (10%), and cross-domain applications (10%). These distributions indicate that digital twin research in computer engineering is concentrated in energy-related CPS but expanding toward industrial, networking, and smart-environment contexts.

Publication type and publication outlet

Among the 45 articles reviewed for this study, almost all were published as full journal articles. Based on the recorded venues, DOIs, and links, 44 articles (≈98%) appeared in peer-reviewed journals, while one article (≈2%) was published in an outlet that could not be clearly classified but was retained because it met all other inclusion criteria. No clearly identifiable conference proceedings remained in the final sample after applying the exclusion criteria for work-in-progress and short papers.

The journal articles were distributed across a wide range of outlets in computer and electrical engineering, networking, and cyber-physical systems. The most common outlet was Future Internet (3 articles, ≈7% of all journal papers). Several venues each published two of the sampled articles (≈4.5% each): IET Cyber-Physical Systems: Theory and Applications, IEEE Access, Sensors, Future Generation Computer Systems, Advanced Science, Digital Twins and Applications, Advanced Engineering Informatics, and two journals recorded generically under Elsevier in the data extraction sheet. The remaining journals each contributed a single article (≈2–3% each), including IEEE Internet of Things Journal, IEEE Wireless Communications Letters, IEEE Transactions on Smart Grid, IEEE Transactions on Industrial Informatics, IEEE Open Journal of the Communications Society, IEEE Network, IET Generation, Transmission and Distribution, International Journal of Electrical Power and Energy Systems, Renewable and Sustainable Energy Reviews, Applied Energy, Journal of Electrical and Computer Engineering, Software: Practice and Experience, International Journal of Intelligent Systems, Journal of Industrial Information Integration, Advanced Intelligent Systems, Energies, Wireless Personal Communications, IEICE, and Advances in Electrical and Computer Engineering.

Overall, the dominance of archival journals and the concentration in a set of high-impact computer and electrical engineering outlets suggest that research on digital twin technology in this domain is being positioned as mature, archival scholarship rather than preliminary conference work.

Country affiliation of first author

Table 3 shows that the articles selected for this review featured first authors from 18 countries, plus two articles for which the country could not be uniquely identified and were coded as “Not specified.” The largest shares of first authors hailed from the United States and China (each 6 articles, 13.3% of the sample), followed by India (5 articles, 11.1%). First authors from the United Kingdom contributed 4 articles (8.9%), while South Korea and Japan each accounted for 3 articles (6.7% each). Countries with 2 first-author contributions (4.4% each) included Italy, Singapore, Brazil, and Greece. The remaining countries, each represented by a single article (2.2% each), were Finland, Tunisia, Denmark, Egypt, Romania, Spain, Canada, and Nigeria, along with two articles coded as Not Specified because the author list reflected multiple or international affiliations that could not be assigned to a single country.

TABLE 3. Distribution of sampled articles by first author country.

This distribution suggests that current digital twin research in computer and electrical engineering is concentrated in a relatively small set of technologically advanced regions in North America, Europe, and Asia, while still drawing contributions from a diverse group of countries. At the same time, the predominance of English-language, open-access articles in our inclusion criteria likely amplifies the visibility of work from these regions and may underrepresent scholarship from countries where publishing practices or access models differ.

Application domains and areas represented in sampled articles

Table 4 lists the main application domains and application areas represented in the 20 articles that included a clearly specified digital twin use case. Across all domains, power and energy systems were the most frequently studied (55.0%), including work on smart grids, networked microgrids, renewable-dominant power systems, substations, lithium-ion batteries, and DC microgrids. The second most common domain was Industry 4.0 and manufacturing or maintenance (20.0%), where digital twins were used for smart factories, heavy machinery such as loaders, and predictive or structural-health maintenance scenarios. A smaller but still notable subset of studies focused on networking, extended-reality, and smart retail applications (10.0%), using digital twins to support 5G/IoT sensing, VR/AR experiences, and AIoT-enabled virtual shopping environments.

TABLE 4. Application domains and areas represented in sampled articles.

Percentages are calculated with respect to the 20 articles that reported a specific application domain.

In addition, several articles addressed cross-domain or multi-sector applications (10.0%), providing comparative analyses or surveys of digital twin use across multiple engineering contexts rather than a single system. Only one application paper (5.0%) centered on laboratory and smart-environment settings, using digital twins to monitor and detect anomalies in IoT-driven smart labs. These findings show that, while digital twins are being explored in a variety of cyber-physical domains, current work is heavily weighted toward energy-related systems, with growing but still limited coverage of industrial, networking, and smart-environment applications. (Note: the remaining 25 articles in the sample were not included in Table 3 because they primarily comprised frameworks, architectures, or broad surveys where a specific application domain was not indicated).

Theoretical frameworks

Across the forty-five articles included in this review, a wide range of theoretical and conceptual frameworks were used to guide model development, system architecture, simulation, and evaluation (please refer to Table 5). Eighteen articles employed an explicit theoretical or architectural framework grounded in existing literature or established methodologies. Eleven articles used a conceptual framework derived from domain-specific models such as Cyber-Physical Systems (CPS), Digital Twin (DT) reference architectures, multi-agent systems, or Model-Based Systems Engineering approaches. The remaining fifteen articles did not use a formal framework, instead focusing on algorithmic development, application prototypes, or engineering implementations without anchoring them in a named theoretical model.

TABLE 5. Frameworks used in the sampled articles.

Among the papers that used an explicit framework, the diversity of approaches reflects the multidisciplinary nature of digital twin research. Frameworks ranged from systems engineering, learning-based control, agent-based modeling, IoT-cloud architectures, reinforcement learning theory, edge computing optimization models, and structural simulation frameworks. No single framework appeared repeatedly across multiple articles, indicating that the field has not yet converged on a unified theoretical perspective for digital twin design. However, cross-cutting themes such as real-time synchronization, computational modeling, prediction-based control, and system-level optimization appeared frequently in the conceptual structures authors used.

Articles that did not employ a formal framework generally presented new DT architectures, proposed pipelines, or domain-specific implementations, such as renewable-energy DTs, manufacturing DTs, healthcare analytics DTs, microgrid co-simulation environments, or battery management twins. These studies tended to emphasize engineering validation rather than theory-driven analysis.

The fifteen articles without a theoretical framework mainly focused on applied engineering solutions where the contribution involved system design rather than theoretical grounding. These included cloud-edge orchestration models, task-offloading strategies, reinforcement-learning-enhanced DTs, structural fault detection pipelines, and sustainability-driven digital twin deployments. Overall, the diversity of frameworks suggests an active and evolving research space with multiple perspectives informing the development of digital twin technologies across engineering domains.

Research foci and research methods

Table 6 summarizes the different research foci and methods represented in the forty-four selected Digital Twin (DT) and Cyber-Physical System (CPS) articles. Five broad categories of research emerged from the dataset: literature reviews (11.4%), framework or system architecture descriptions (34.1%), intervention or implementation studies (31.8%), domain-specific application studies (9.1%), and fundamental or theoretical model development (13.6%).

TABLE 6. Research foci and research approach in sampled digital-twin articles.

Studies coded as “no research (descriptive)” primarily proposed architectures, workflows, or DT pipelines without collecting new data. These descriptive works typically detailed system structures, data-flow mechanisms, and functional models across manufacturing, energy, and healthcare domains. In contrast, studies applying qualitative, quantitative, or mixed methods collected and analyzed data related to model validation, simulation accuracy, energy efficiency, or human-machine interaction.

The fifteen papers in the architecture description category focused on the design or documentation of DT frameworks, CPS-IoT integrations, and edge-cloud orchestration systems. Among these, four adopted quantitative performance analyses (e.g., latency or accuracy benchmarking), while five combined simulation data and algorithmic evaluation in a mixed-method format. The fourteen studies categorized under intervention or implementation used quantitative or mixed approaches to assess the impact of DT deployment in specific settings such as industrial robotics, renewable-energy microgrids, or battery-health prediction. These studies typically included (1) performance data from sensors or emulated physical systems, (2) comparative simulations between baseline and twin-enabled models, and (3) validation metrics like prediction error or energy savings. Nine of these intervention studies applied inferential statistics to evaluate framework efficiency or fault-detection reliability.

Six papers were coded as fundamental research, emphasizing theoretical or mathematical contributions such as physics-based simulation modeling, model-based systems engineering, or reinforcement-learning-driven DT optimization. All studies in this category employed quantitative or analytical modeling methods and demonstrated cross-domain applicability beyond individual use-cases. Finally, four studies were domain-application papers, mainly describing DTs for smart agriculture, circular economy monitoring, and security analysis. These were descriptive or qualitative, detailing data-integration challenges and technical feasibility.

The fifteen frameworks or architecture papers predominantly proposed DT reference architectures or data-collection frameworks for CPS environments, including IoT-edge coordination (Zhang et al., 2022) and AIoT-based virtualization systems (Sun et al., 2021). The fourteen intervention studies mainly implemented DT solutions experimentally. Examples include real-time co-simulation for microgrids (Cao et al., 2019), federated learning for turbine monitoring (Ribeiro et al., 2025), and reinforcement-learning-driven control for energy optimization (Seyedi et al., 2025). Six papers were classed as fundamental research, exploring digital-twin theory and algorithmic foundations such as MBSE (Zhang et al., 2023), physics-based modeling (Zhang et al., 2023), and hybrid AI-DT learning systems (Kerkeni et al., 2025). These studies advanced theoretical understanding rather than system-specific application. The five literature reviews covered topics such as artificial-intelligence integration into DTs, security architectures for cyber-physical systems, sustainability frameworks for industrial DTs, and predictive maintenance models for manufacturing and energy infrastructure (e.g., Zhang et al., 2023; Doenhoff et al., 2022; Nair et al., 2024). Finally, four domain-application studies described DT deployment in smart irrigation, security monitoring, or agriculture automation contexts (Arnaoutoglou et al., 2025; Doenhoff et al., 2022).

Thematic analysis: descriptions and implications

Using inductive thematic analysis, 45 articles were organized into six themes (Table 7). Articles could map to more than one theme when their contributions span multiple areas.

TABLE 7. Distribution of sampled articles based on thematic classification.

Figure 3 displays the frequency of occurrence of each theme across the included digital-twin publications from 2017 to 2026. Modeling, simulation, and testing of digital twin systems (Theme 5) and application domains and use cases (Theme 6) become increasingly prominent in the later years, while work on intelligent digital twins (Theme 3) and networked or distributed infrastructures (Theme 2) also grows steadily, indicating a shift from foundational concepts toward data-driven, deployed cyber-physical systems.

To clarify the boundaries among the six themes, the thematic synthesis treats each theme as emphasizing a different level of the digital twin stack. Theme one focuses on the local computer-engineering foundations of digital twin systems, such as embedded platforms, hardware-software co-design, and real-time execution. Theme 2 focuses on communication and distributed infrastructure across IoT, edge, cloud, and future wireless environments. Theme 5 focuses on modeling, simulation, interoperability, and validation practices rather than deployment infrastructure itself. This distinction reduces overlap among discussions of edge-cloud orchestration, co-simulation, and CPS architecture while preserving the cross-theme connections that are inherent to deployed digital twin systems.

Theme 1: digital twin foundations, architecture, and delivery for computer engineering

Nine articles addressed how digital twin (DT) capabilities are designed, implemented, and operationalized for computer engineering tasks, ranging from foundational surveys to concrete architectures for cyber-physical systems (CPS), communications infrastructure, and AI-driven control (please see Table 8). This theme is limited to the local architectural and runtime foundations that determine whether a twin can execute reliably on real computing platforms. Collectively, these studies show that effective DT implementation depends on end-to-end system design rather than surface-level visualization: data pipelines, timing and synchronization, interoperability, compute placement between edge and cloud resources, and validation against real systems. For example, Adekola et al. (2024) synthesize DT as a practical tool across computer engineering work, emphasizing DT’s role in mirroring real assets and supporting monitoring, optimization, and decision-making. We categorized four articles under this theme as DT-enabling platform and infrastructure work, including (i) a broad DT-in-computer-engineering review (Adekola et al., 2024), (ii) a WoT-based architecture for a Relativistic Digital Twin (Sciullo et al., 2023), (iii) a framework connecting on-device AI with DTs as a joint approach for AI development and deployment (Barbone et al., 2025), and (iv) a hardware-focused perspective on next-generation wireless communication infrastructure needed for the Future Internet, which directly constrains DT connectivity and performance (Arnaoutoglou et al., 2025).

TABLE 8. Artiles categorized in theme 1: computer engineering foundations of digital twin systems.

We categorized an additional four articles as DT-driven CPS design and real-time control engineering, where the DT is treated as a predeployment verification and risk-reduction instrument. Dai and Burns (2020) propose a runtime adaptation method for long-lived CPS that uses online monitoring and model-based prediction to safely adjust control-task periods under fixed-priority scheduling, explicitly leveraging DT ideas to manage performance degradation within acceptable bounds. Cao et al. (2019) present a real-time CPS co-simulation testbed for microgrids, coupling power and communication simulators and providing an environment to verify control algorithms and study cyber-event impacts. Ahmed et al. (2024) propose a multi-agent-system-based CPS model with nine layers (spanning physical through application concerns) to improve CPS flexibility and scalability. Finally, Seyedi et al. (2025) evaluate DT-enabled transfer learning for reinforcement-learning control in DC microgrids, using the DT for safer training and then testing on an actual testbed, explicitly highlighting reduced risk versus training directly on the system. We categorized one additional article as DT-enabled application delivery in an AIoT setting: Sun et al. (2021) demonstrate AIoT-enabled “virtual shop” applications using a sensor-enhanced soft robotic manipulator, showing how DT-aligned sensing and intelligent perception can support interactive, real-world deployment scenarios.

Exemplar studies

Exemplar studies under this theme include two papers that connect DT architecture decisions to operational constraints. First, Sciullo et al. (2023) introduce a Relativistic Digital Twin architecture via “Digital Twin Web Thing,” aiming to integrate DTs into the Web of Things ecosystem and enable predictive “what-if” style reasoning while managing latency and consistency issues. Second, Seyedi et al. (2025) use a DT as a training ground for RL-based control in DC microgrids, arguing that DT-based training mitigates interruptions and safety concerns from exploratory actions during on-system RL training, then benchmarking against PI control on a real testbed. A third strong exemplar is Cao et al. (2019), who build a real-time co-simulation testbed that explicitly models the coupling between energy flow and information flow, enabling experiments on how cyber events influence microgrid control performance.

Research implications

Future DT research in computer engineering should more closely integrate between (1) interoperable architectures (for example, Web-of-Things integration) and (2) hard real-time constraints like synchronization, scheduling jitter, and cross-platform co-simulation timing. There is also a clear need for rigorous evaluation methods for DT-to-physical-system transfer, including standardized metrics for fidelity, robustness under cyber disturbances, and safety-aware learning workflows for AI controllers trained in DT environments. Finally, DT systems are increasingly bottlenecked by where intelligence runs (cloud vs. edge), making on-device AI and communications hardware constraints first-class DT design variables rather than afterthoughts.

Practice implications

For practitioners, these studies imply four actionable takeaways. First, treat interface design and synchronization as core requirements (rather than deferred integration work) when combining power, control, and network simulators in DT/co-simulation pipelines. Second, use DTs to de-risk deployment of adaptive control and learning-based controllers, especially when exploratory actions can cause instability or downtime. Third, structure CPS DT solutions with explicit layers and agent responsibilities when scalability, coordination, and maintainability are key, as MAS-style decompositions can make complex CPS easier to evolve. Fourth, engineer DT delivery around real compute and network limits, meaning: plan for on-device intelligence where latency or connectivity is tight, and align the DT data/communications design with realistic wireless hardware capabilities.

Theme 2: networked and distributed infrastructure for digital twins

We categorized fifteen articles under networked and distributed infrastructure for digital twins. This theme is distinguished from Theme one by its focus on inter-node communication, orchestration, and distributed computation rather than local platform execution (please refer to Table 9). All articles underscored the importance of advanced networking and distributed computing resources in enabling scalable, real-time digital twin systems and support the notion that digital twins achieve higher fidelity, responsiveness, and operational value when they are tightly integrated with wireless communication networks, IoT sensing infrastructures, and edge-cloud computing environments (Khan et al., 2022a; Khan et al., 2022b; Tang et al., 2022; Gupta et al., 2023; Zhang et al., 2022; Tao et al., 2023; Suganya et al., 2024; Karobi et al., 2025; Arnaoutoglou et al., 2025; Do-Duy et al., 2022; Mitsiou et al., 2023; Zhou et al., 2022; Saad et al., 2020a; Saad et al., 2020b; Salehiyan et al., 2025).

TABLE 9. Articles categorized in theme 2: networked and distributed infrastructure for digital twins.

Exemplar studies

We chose the exemplar studies for this theme because they centered on the role of networking and distributed computation in enabling effective digital twin operation, with a particular focus on edge–cloud coordination, orchestration, and real-time decision-making. Tang et al. (2022), an exemplar of this theme, introduced the concept of Digital Twin Edge Networks, which integrate digital twins with mobile edge computing to support centralized intelligence over distributed wireless environments. Their work outlined architectural components for DT modeling, deployment, and updating, and demonstrated how DT-enabled edge networks can improve system-level optimization while operating under latency, bandwidth, and resource constraints common in 5G and emerging 6G systems.

Another exemplar, Khan et al. (2022a), conceptualized digital-twin-enabled 6G as a foundational framework for future wireless networks and Internet of Everything services. They argued that digital twins will function as core components for intelligent network management, proactive optimization, and real-time adaptation, enabled by the convergence of AI, edge computing, and ultra-reliable low-latency communications. In a related exemplar, Khan et al. (2022b) developed a taxonomy of digital twins of wireless systems, distinguishing between digital twins designed to optimize wireless networks and wireless technologies designed to support digital twin applications, thereby emphasizing the bidirectional relationship between communication infrastructure and digital twin performance.

Research implications

Potential directions for future work include investigating how digital twin accuracy and responsiveness are affected by imperfect, delayed, or incomplete data in distributed network environments; determining optimal strategies for partitioning digital twin computation across edge and cloud resources to balance latency, scalability, and energy consumption; and examining how secure and privacy-preserving learning mechanisms can support continuous digital twin updates in large-scale deployments (Tang et al., 2022; Zhou et al., 2022; Gupta et al., 2023). Further, despite the growing importance of networked digital twins, relatively few studies have validated proposed architectures under real-world deployment conditions, as many evaluations rely on simulations or controlled prototypes. This highlights the need for longitudinal, large-scale experimental studies that assess robustness, interoperability, and sustainability across heterogeneous infrastructures (Arnaoutoglou et al., 2025; Karobi et al., 2025).

A further gap concerns blockchain and distributed ledger technologies for digital twin coordination. Although the reviewed studies emphasize distributed architectures, real-time synchronization, and interoperability, they rarely examine whether ledger-based mechanisms could support decentralized trust, tamper-evident data exchange, provenance tracking, and cross-organization auditability in multi-stakeholder DT deployments. Recent work on DT-DLT convergence and blockchain-based access management suggests that distributed ledgers can strengthen trust and interoperability, but they also introduce latency, throughput, storage, and governance trade-offs that may conflict with real-time CPS requirements (Hemdan et al., 2023; Somma et al., 2024; Ali et al., 2025). Future research should therefore evaluate when blockchain improves DT coordination and when lighter-weight trust or access-control mechanisms are more appropriate.

Practice implications

Articles under this theme highlighted that designers of digital twin systems should intentionally integrate networking and distributed computing considerations into system architecture, as these factors play a critical role in determining performance, security, and scalability (Khan et al., 2022a; Tang et al., 2022). The studies under this theme provide several practical strategies for improving digital twin deployment. For example, practitioners can leverage edge computing to support latency-sensitive digital twin functions while using cloud resources for computation-intensive analytics and long-term model management (Do-Duy et al., 2022; Mitsiou et al., 2023). Secure and latency-aware scheduling mechanisms can be incorporated to improve reliability in safety-critical applications such as power grids and industrial control systems (Zhou et al., 2022; Saad et al., 2020b). Additionally, adopting standardized reference architectures can improve interoperability and system integration across heterogeneous platforms (Tao et al., 2023). Finally, integrating intelligent sensing, intrusion detection, and sustainability-aware design principles can enhance the resilience, security, and long-term viability of digital twin infrastructures as they evolve toward 6G-enabled environments (Zhang et al., 2022; Suganya et al., 2024; Salehiyan et al., 2025; Karobi et al., 2025).

Theme 3: computer engineering for intelligent digital twins

We categorized fifteen articles under computer engineering approaches that make digital twins (DTs) intelligent, meaning the twin is not only a synchronized replica, but also learns patterns, detects anomalies, makes predictions under uncertainty, and supports real-time decision making (please refer to Table 10). Collectively, these studies show DT intelligence is being built through (1) AI/ML model integration for prediction and classification, (2) trust, uncertainty, and probabilistic representations for “safer” cyber-physical interaction, and (3) deployment-ready control and perception pipelines that work under real-world constraints like limited data, edge compute limits, and noisy sensing (Saraf and Pachghare, 2025; Shimonishi et al., 2024; Ribeiro et al., 2025; Antonesi et al., 2025; Barbone et al., 2026; Salehiyan et al., 2025; Gupta et al., 2024; Wu et al., 2025; Kerkeni et al., 2025; Xin et al., 2023; Nair et al., 2024; Seyedi et al., 2025; Zhang et al., 2022; Sun et al., 2021; Karobi et al., 2025).

TABLE 10. Articles categorized in theme 3: computer engineering for intelligent digital twins.

Exemplar studies

We selected exemplar studies for this theme because they directly operationalize “intelligence” in a DT: not just presenting a concept, but showing how learning, prediction, or control is actually implemented and evaluated. Nair et al. (2024) is an exemplar because it builds a discharge-capacity DT for Li-ion batteries by pairing a semiempirical capacity model with ML models, then explicitly optimizing ML hyperparameters using metaheuristics (e.g., IGWO, GWO, antlion) and validating with ten-fold cross validation on the NASA battery aging dataset. The work reports very low error for the best configuration (IGWO-AdaBoost DT) and frames the DT as a practical real-time predictor when direct capacity determination is difficult. Another exemplar is Seyedi et al. (2025), which uses a DT as a safe training platform for reinforcement-learning control in a DC microgrid, explicitly arguing that training on the real system can cause interruptions and safety risks. Their DT-trained RL controller is then tested on an actual testbed and compared against a PI baseline, emphasizing the DT’s role in reducing risk during learning, and they report stronger disturbance responses and performance improvements under fault scenarios in the experimental results.

A third exemplar angle is “DT intelligence through edge-native optimization.” Karobi et al. (2025) presents EcoEdgeTwin, positioning intelligence as predictive and adaptive edge control for 6G-era networks by optimizing a utility function that balances QoE, latency, and energy, and by operationalizing mechanisms like task offloading, caching, and service migration, with deep reinforcement learning (advantage actor-critic) used for dynamic management. This matters because it treats the DT as an active decision engine, not a passive mirror.

Research implications

Across Theme 3, a major research direction is closing the gap between “laboratory-grade models” and deployable DT intelligence. One thread is improving learning under limited, shifting, or uncertain data. For example, Sun et al. (2021) shows how multimodal self-powered sensing plus ML can classify objects (reported high recognition accuracy across many object shapes) and feed a DT-based “virtual shop” concept, but it also highlights how richer attributes (like freshness) demand additional sensing modalities and data fusion decisions. Another thread is robust DT control and decision making: Seyedi et al. (2025) indicates that DT-enabled RL can address training safety and sim-to-real challenges, but it also implies future work should generalize to more complex microgrid topologies and conditions. On the systems side, Zhang et al. (2022) frames DT intelligence as emerging from sensor-rich data pipelines in 5G/IoT contexts, and explicitly points to power constraints as a bottleneck that motivates self-powered sensing (e.g., TENG-based systems) and sensor fusion as a next-step capability for DT ecosystems. Finally, EcoEdgeTwin’s framing suggests a research gap around how DT estimation error impacts downstream policies, because performance may degrade nonlinearly as DT prediction error grows.

In parallel, trustable DT intelligence remains an open research problem. Papers in this theme explicitly push trust computation (Saraf and Pachghare, 2025), probabilistic DT representations for trustable cyber-physical interactions (Shimonishi et al., 2024), and probabilistic modeling for spatio-temporal DTs (Ribeiro et al., 2025), which collectively suggests a convergence: intelligent DTs will increasingly need explicit uncertainty and trust signals rather than only point predictions (Antonesi et al., 2025; Barbone et al., 2026).

Practice implications

Practically, these studies show that a DT designed for real-world decision support must account for (1) sensing and data quality, (2) deployable learning and control, and (3) measurable reliability. For sensing-heavy DT applications, Zhang et al. (2022) and Sun et al. (2021) together imply that engineers should treat sensing choice, energy constraints, and sensor fusion as core DT architecture decisions rather than secondary implementation details. For control-heavy DT applications, Seyedi et al. (2025) shows a defensible workflow: train RL in the DT to avoid unsafe exploration on the real system, then validate on a physical testbed against a baseline controller. For predictive-maintenance and reliability tasks, battery DT work such as Nair et al. (2024) suggests combining physics-grounded structure, such as semiempirical capacity estimation, with ML correction and rigorous cross-validation. For edge/6G DT infrastructures, EcoEdgeTwin’s approach implies that engineers should explicitly encode QoE, energy, and latency trade-offs into the optimization objective and treat mobility, caching, and offloading as first-class control variables. Overall, intelligent DTs are full-stack systems, so weak sensing, modeling, networking, or validation at any layer can reduce reliability in cyber-physical operation.

Theme 4: security, privacy, and dependability in DT-enabled CPS

We categorized eight articles under security, privacy, and dependability in digital twin (DT) enabled cyber-physical systems (CPS) (please see Table 11). Collectively, these studies argue that DTs are not merely passive replicas of physical assets; they can function as active security infrastructure: modeling attacks, quantifying uncertainty, computing trust, and enabling responsive mitigation across edge, cloud, and operational technology (OT) environments (Saad et al., 2020a; Saraf and Pachghare, 2025; Doenhoff et al., 2022; Shimonishi et al., 2024; Suganya et al., 2024; Zhou et al., 2022; Salehiyan et al., 2025; Gupta et al., 2024).

TABLE 11. Articles categorized in theme 4: security, privacy, and dependability in DT-enabled CPS.

Exemplar studies

We chose the exemplar studies in this theme to emphasize how DTs strengthen CPS security via (1) practical mitigation frameworks, (2) uncertainty-aware trustability, and (3) scalable security-data pipelines. Saad et al. (2020a) provides an IoT-based DT framework for networked microgrids resiliency, explicitly designed to respond to multiple cyberattacks, including coordinated false data injection and denial-of-service, and validated on a practical distributed control setup where security measures were implemented via cloud computing and controllers used single-board computers.

Another exemplar, Shimonishi et al. (2024), argues that “trustable CPS” fundamentally requires probabilistic digital twins, where objects’ identities and states are stored as probability distributions and linked via probabilistic graphical structures to represent spatial, temporal, and causal dependencies, enabling risk-sensitive prediction and control under uncertainty. A third exemplar, Doenhoff et al. (2022), targets a major real-world barrier: security DTs fail in practice if data collection is messy and expensive. They propose a data-collection abstraction layer for a modular Security Digital Twin (SDT), using a unified interface driven by upper-layer requirements, and qualitatively show it improves cost-efficiency in OT/IoT contexts.

Beyond exemplars, Saraf and Pachghare (2025) operationalize security and trust selection in federated cloud computing by proposing DTUPTC, where users specify preferred QoS attributes and desired values and the model dynamically computes weights without manual intervention, positioning DTs as a mechanism for evidence-based trust computation under security and non-functional constraints. Saad et al. (2020a) further reinforces the “DT as defender” view by showing a DT can participate directly in coordinated-attack mitigation while maintaining CPS control-loop integrity.

Research implications

Across these articles, three research gaps recur. First, DT security needs stronger closed-loop integration between detection, simulation, and actuation so that the DT does not merely diagnose problems but helps decide and deploy mitigations in near real-time (Saad et al., 2020b). Second, “trustability” research needs to treat uncertainty as a first-class citizen; probabilistic DT representations suggest a path, but generalizable validation across more CPS domains is still needed (Shimonishi et al., 2024). Third, scaling security DTs depends on data plumbing: abstraction layers and standardized interfaces are essential, especially for heterogeneous OT/IoT device fleets where naive collection becomes cost-inefficient (Doenhoff et al., 2022).

In addition, the security literature reviewed here gives limited attention to blockchain-enabled or distributed-ledger-based DT architectures. This omission matters because several security and dependability problems identified in this theme, including data integrity, identity management, auditability, and trust across organizational boundaries, align closely with the capabilities often associated with blockchain-based access control and tamper-evident logging. At the same time, blockchain should not be treated as a default solution for all DT security problems; consensus delay, transaction cost, storage overhead, and privacy leakage can undermine latency-sensitive CPS applications. A useful research direction is therefore to compare blockchain-supported trust architectures with conventional centralized, federated, and zero-trust mechanisms under shared DT benchmarks.

Practice implications

For computer engineering practice, these studies imply DT systems should be architected as security-capable CPS components, not passive replicas. Practitioners can (1) implement DT-integrated resilience mechanisms for CPS such as microgrids, including testing DT behavior against coordinated attack classes (Saad et al., 2020b), (2) adopt probabilistic modeling and risk-sensitive controls when sensor noise, identity ambiguity, and network instability make deterministic DTs unsafe (Shimonishi et al., 2024), and (3) design modular SDTs with unified data-collection interfaces to reduce engineering overhead and enable reusability across OT/IoT deployments (Doenhoff et al., 2022). Finally, trust and service selection in DT-enabled cloud/edge stacks can be made more aligned with user requirements by incorporating user-preference QoS and DT-simulated evidence into trust scoring (Saraf and Pachghare, 2025).

Theme 5: modeling, simulation, and testing of digital twin systems

We categorized twenty-five articles (please see Table 12) under modeling, simulation, and testing because they collectively focus on how computer engineering teams build, evaluate, and validate digital twins: defining reference architectures, selecting modeling approaches (physics based, data-driven, hybrid), enabling co-simulation/testbeds, and addressing interoperability across cyber-physical systems (CPS) and domains (Adekola et al., 2024; Acharya et al., 2024; Alam and El Saddik, 2017; Antonesi et al., 2025; Cao et al., 2019; Dai and Burns, 2020; da Silva Mendonça et al., 2022; Fuller et al., 2020; Gil et al., 2024; Han et al., 2026; Khan et al., 2022a; Khan et al., 2022b; Karobi et al., 2025; Mansour et al., 2023; Rocha-Jácome et al., 2022; Sciullo et al., 2024; Somers et al., 2023; Tang et al., 2022; Vidyalakshmi et al., 2025; Wu et al., 2021; Xiong et al., 2025; Zhao et al., 2025; Doenhoff et al., 2022; Ahmed et al., 2024; Zhang et al., 2022). This theme is distinguished from Theme 2 by its emphasis on model credibility, test design, validation, and interoperability evidence rather than the networking infrastructure that transports DT data. Across these works, the theme is consistent: DTs become engineering tools only when their models, data pipelines, simulation links, and validation routines are specified well enough to be repeatable, testable, and interoperable across hardware, networks, and software stacks.

TABLE 12. Articles categorized in theme 5: modeling, simulation, and testing of digital twin systems.

Exemplar studies

We chose exemplar studies that most directly operationalize this theme from a computer engineering perspective: (1) reference modeling and system decomposition, (2) rigorous testing/validation framing, and (3) practical framework/testbed guidance. First, Alam and El Saddik (2017) introduced C2PS, a digital twin architecture reference model that explicitly maps the physical system into cyber computation modules, clarifying how sensing, data handling, and compute services should be organized so that DT implementations are not one-time prototypes but reusable CPS building blocks. Second, Somers et al. (2023) framed DTs as a testing asset, not just a monitoring dashboard, by systematically reviewing DT based testing for CPS and emphasizing core test engineering questions like how test oracles are defined, what modeling visibility is assumed (white/black/grey box), and how test cases and inputs are generated when the twin is coupled to a running system. Third, Gil et al. (2024) tackled the “implementation feasibility” problem by surveying open source DT frameworks via case study style comparison, highlighting that DT implementations require multiple components working together (modeling/simulation, data ingestion, synchronization, interfaces), and that framework choice directly shapes feasibility for prototyping, experimentation, and reproducible evaluation in computer engineering research.

Beyond the exemplars, many studies reinforce the same iterative build-test-evaluate pipeline but from different angles. Communications and networking focused works treat the DT as a networked system artifact: Khan et al. (2022b) positioned DTs as a foundational capability for 6G era systems, where distributed intelligence, edge/cloud resources, and strict latency requirements push DT architectures toward layered designs that can keep a synchronized “live” cyber replica. Complementing that, Khan et al. (2022a) surveyed DTs for wireless systems with a taxonomy of requirements, challenges, and opportunities, which matters for computer engineers because wireless DTs are difficult to validate if timing, data fidelity, and resource constraints are not handled as first class design variables. Relatedly, Tang et al. (2022) surveyed Digital Twin Edge Networks (DITEN) toward 6G, pushing the idea that DT computation and coordination must be edge aware (offloading, placement, and orchestration become part of the DT design space). At a higher abstraction level, Wu et al. (2021) defined the DTN (digital twin network) concept and summarized enabling technologies and open issues, underscoring that DTs are not just models but also require networked communication, computing, and management mechanisms. Sciullo et al. (2024) extended this “network reality” focus by emphasizing IoT scale considerations for DTs, which in practice pulls computer engineering attention toward standardized descriptions, time alignment, and dependable data-to-model synchronization.

Power and energy-oriented surveys add another crucial computer engineering angle: DTs become credible when they are connected to operational constraints, grid data streams, and validation environments. Mansour et al. (2023) surveyed IoT and DT in electrical power systems, organizing applications and challenges in a way that highlights integration burdens, including data pipelines, connectivity, and system-level constraints that are central to CE workflows. Han et al. (2026) reviewed DT technologies for active distribution networks, where distributed energy resources and grid dynamics increase the need for structured models and validated simulation workflows. Zhao et al. (2025), via a guest editorial on renewable-dominant power systems, emphasized methodologies and application directions that recur across modern DT work, including tighter integration between modeling, simulation, and operational decision-making. Moving into next-generation AI-enabled DTs, Antonesi et al. (2025) surveyed transformers and large language models in the energy sector and connected these capabilities to agentic digital twins, an emerging direction in which DTs not only simulate systems but also support planning, reasoning, and action when verification and guardrails are carefully engineered.

Several works in this theme exist to prevent DT research from collapsing into incompatible silos. Acharya et al. (2024) directly addressed interoperability levels and challenges for DTs in CPS, which matters because CE DT stacks often span heterogeneous sensors, protocols, simulation tools, and data models. Xiong et al. (2025) took a cross domain comparative approach, arguing that sector specific DT solutions keep reinventing the wheel and motivating more universalized DT solutions, including shared platform patterns and interoperability strategies. Fuller et al. (2020) similarly cataloged enabling technologies and open research challenges, which helps explain why interoperability remains hard: DTs sit at the intersection of sensing, comms, compute, modeling, and visualization, so weak integration in any one layer ruins end to end reliability. Complementing these framework oriented surveys, Zhang et al. (2022) provided a broad review of recent digital twin applications across multiple engineering domains and explicitly highlighted that differences in modeling fidelity, simulation depth, and validation maturity remain a key barrier to transferring digital twin solutions across domains, reinforcing the need for standardized testing methodologies and interoperable DT architectures in computer engineering contexts.

The remaining studies strengthen the “modeling-simulation-testing” story by broadening application coverage and emphasizing repeatability. da Silva Mendonça et al. (2022) surveyed DT applications and challenges across domains, reinforcing recurring constraints like real-time data integration, scaling, and security, which repeatedly show up as testability blockers. Vidyalakshmi et al. (2025) positioned DTs and CPS as a new frontier in computer modeling, reinforcing that DTs are best treated as an engineering paradigm rather than a single tool, since real deployments require integrating models, data, and feedback loops. At the operational layer, Dai and Burns (2020) contributed a control scheduling-oriented perspective by proposing runtime period adaptation for CPS control tasks using monitoring and prediction, which highlights a core CE reality: DT adjacent systems have to respect real-time scheduling and resource budgets. Finally, data acquisition is part of DT “modeling” because model validity depends on data quality: Doenhoff et al. (2022) focused on data collection methods for security digital twins in CPS, reinforcing that without consistent, structured data capture, the twin’s model fidelity and test conclusions are difficult to trust.

Research implications

Future work should push harder on comparability and falsifiability of DT studies. First, DT research needs clearer reporting standards for model fidelity, synchronization strategy, and evaluation protocols so that surveys and frameworks can be tested against each other instead of just summarized (Somers et al., 2023; Gil et al., 2024; da Silva Mendonça et al., 2022). Second, interoperability should be treated as a measurable engineering target (levels, semantics, interfaces), not a vague limitation section, because cross tool and cross vendor DT stacks are the norm in computer engineering CPS (Acharya et al., 2024; Xiong et al., 2025; Alam and El Saddik, 2017). Third, as AI heavy DTs become more common, research has to define validation boundaries: what parts are simulated, what parts are learned, what failure modes are acceptable, and what guarantees exist when DT outputs influence real systems (Antonesi et al., 2025; Khan et al., 2022b; Mansour et al., 2023). Finally, more DT work should publish reusable artifacts: testbeds, reference implementations, and benchmark datasets that reflect timing, network, and resource constraints typical in CE deployments (Cao et al., 2019; Tang et al., 2022; Zhao et al., 2025).

Practice implications

For DT systems in computer engineering contexts, these papers collectively point to several core requirements. Researchers and practitioners should begin with an explicit architecture or reference model so the DT is engineered as an integrated system rather than a collection of disconnected scripts (Alam and El Saddik, 2017; Fuller et al., 2020). Modeling and simulation approaches should align with the operational constraints of the target CPS, including real-time scheduling, networking limits, and data availability (Dai and Burns, 2020; Khan et al., 2022a; Wu et al., 2021). Testing and validation should be treated as first-class concerns, including defining test oracles, expected behaviors, and coupling assumptions (Somers et al., 2023; da Silva Mendonça et al., 2022). Interoperability should be built into the system early through structured data models and interfaces, because retrofitting semantics and integration mechanisms later can create avoidable engineering risk (Acharya et al., 2024; Doenhoff et al., 2022; Xiong et al., 2025). Finally, when DTs affect critical infrastructure such as power systems, robust data pipelines, cybersecurity awareness, and credible evaluation environments such as co-simulation testbeds should be prioritized (Mansour et al., 2023; Han et al., 2026; Cao et al., 2019).

Comparative validation metrics and benchmarking practices

Across the reviewed studies, validation practices vary substantially by domain and maturity level. Simulation-based and architecture-focused papers tend to report model behavior, feasibility, or system-level design properties, while implementation-oriented studies more often report latency, prediction error, resource use, or robustness under disturbances. However, few studies report a complete benchmarking profile that combines timing, fidelity, scalability, dependability, and real-world transfer. Table 13 summarizes the main validation metrics and benchmarking practices synthesized across the reviewed literature.

TABLE 13. Common validation and benchmarking metrics used or needed across digital twin studies.

Taken together, these metrics show that validation in DT, research is not one-dimensional. A DT, that performs well on prediction error may still fail if synchronization lag, energy use, or resilience under cyber-physical disturbance is poor. Therefore, future benchmarking should report both model-level metrics and system-level metrics, with explicit baselines and deployment conditions.

Theme 6: application domains and use cases for digital twins

Theme 6 synthesizes twenty articles that concretely demonstrate where digital twins are being deployed and why those deployments matter in computer engineering, especially when real-time constraints, sensing, cyber-physical coupling, and AI pipelines must be integrated (please see Table 14). Across the set, the dominant application domains were energy systems and power grids, Industry 4.0 operations and maintenance, robotics and human–machine interaction, and immersive or perception-heavy DT interfaces (VR/AR and vision).

TABLE 14. Articles categorized in theme 6: application domains and use cases for digital twins.

Energy systems and grid cyber-physical use cases

Energy-focused papers consistently frame DTs as operational infrastructure for monitoring, optimization, and resilience rather than as visualization layers alone. IoT-based DT architectures are presented for energy cyber-physical systems and for networked microgrids, with an explicit emphasis on design and implementation details and grid resiliency against cyber-attacks (Saad et al., 2020b; Saad et al., 2020a). Power-system scope surveys further position DTs as a unifying layer connecting IoT sensing, analytics, and grid decision-making (Mansour et al., 2023; Ribeiro et al., 2025; Han et al., 2026).

Several studies sharpen this into strict engineering constraints: latency and security-aware scheduling for DT-assisted resource allocation in 5G edge-enabled distribution grids (Zhou et al., 2022), and DT-supported closed-loop control workflows where training directly on physical power hardware is risky or disruptive (Seyedi et al., 2025). A notable pattern here is the rising role of edge and co-simulation: a real-time cyber–physical co-simulation microgrid testbed is proposed to bridge power and communication domains for control studies (Cao et al., 2019). Finally, renewable-dominant power systems are framed as a special case where DT methodology choices materially affect planning and operations, reinforcing that “energy DTs” are not one-size-fits-all (Zhao et al., 2025).

Industry 4.0, maintenance, and asset health use cases

In Industry 4.0 contexts, DTs most often appear as the backbone for predictive maintenance and condition monitoring, with ML serving as the decision engine. One study explicitly combines unsupervised learning with DT concepts for predictive maintenance in Industry 4.0 settings (Kerkeni et al., 2025). Others push into sensor-enabled maintenance demonstrations, like structure-crack detection integrated with a DT demonstration using triboelectric nanogenerators for “intelligent maintenance” workflows (Xin et al., 2023). Battery health is treated as a DT-adjacent forecasting problem, with an AI-driven DT model aimed at reliable lithium-ion discharge capacity prediction (Nair et al., 2024). Outside classical plant-floor assets, DT modeling is also used for heavy equipment performance estimation, where deep transfer learning is paired with DT modeling to predict loader loading resistance (Wu et al., 2025). At a higher organizational level, an Industry 4.0 systematic review proposes ways to organize and scheme Industry 4.0 paradigms from an SLR perspective, which helps explain why DT use cases cluster around integration, automation, and lifecycle digitalization rather than isolated digital models (Rocha-Jácome et al., 2022).

Perception, smart environments, robotics, and VR/AR interaction use cases

Multiple papers show DTs moving toward perception-rich, continuously updated “world models.” For substations, a DT framework is proposed specifically for visual perception under dynamic environmental conditions (Ribeiro et al., 2025). In IoT-driven smart labs, DTs support continuous learning for online anomaly classification, treating the DT as an evolving system that must adapt over time rather than a static replica (Gupta et al., 2024).

Robotics and immersive interaction show up as very “computer engineering” DT use cases: AIoT-enabled virtual shop applications use a DT-based remote interactive system, combining multifunctional self-powered sensing with ML and real-time feedback to users (Sun et al., 2021). More broadly, DTs are positioned as an extension of VR/AR, relying on sensor-driven pipelines that fuse data from the physical world into virtual representations, with a practical definition and mapping emphasis that ties directly to system design (Zhang et al., 2022).

Cross-cutting synthesis: universal patterns, surveys, and “agentic” direction

Several works step back and try to generalize across domains. A broad DT applications survey emphasizes recent advances alongside persistent challenges, reinforcing that deployment challenges (data fidelity, interoperability, lifecycle management) are recurring across use cases (Mylonas et al., 2021). A cross-domain comparative analysis explicitly examines DTs and universalized solutions, supporting the idea that many DT implementations repeat similar architectural motifs even when the application domain differs (Xiong et al., 2025). Finally, a systematic review of transformers and large language models in the energy sector links these models to a trajectory toward “agentic digital twins,” suggesting DTs are evolving from monitoring/optimization tools into more autonomous decision-support systems (Antonesi et al., 2025).

Research implications

Across Theme 6, the research trend is clear: DT value increases when studies tackle real deployment constraints (latency, cybersecurity, co-simulation fidelity, continuous learning drift, and real hardware testing). That shows up directly in latency-aware DT scheduling for edge-enabled distribution grids (Zhou et al., 2022), microgrid cyber-resilience (Saad et al., 2020b), and DT-enabled RL training transferred to an actual testbed (Seyedi et al., 2025). A second implication is that perception-heavy DTs need explicit robustness strategies, not just higher-resolution sensors, as shown in substation visual perception under dynamic conditions and DT-supported anomaly classification that must keep learning online (Ribeiro et al., 2025; Gupta et al., 2024).

Practice implications

Practically, Theme 6 suggests several design requirements for domain-specific DT deployment. First, the DT should be treated as a system architecture decision rather than a dashboard, especially in power and Industry 4.0 contexts where latency, safety, and security are first-class requirements (Mansour et al., 2023; Zhou et al., 2022; Kerkeni et al., 2025). Second, when human interaction is part of the use case, including VR/AR, robotics, or remote operation, the sensing and data-fusion stack determines whether the DT is responsive, trustworthy, and usable (Sun et al., 2021; Zhang et al., 2022). Third, for control and optimization, DTs can reduce real-world risk by acting as training and validation environments before deployment, which is especially relevant when AI controllers are involved (Seyedi et al., 2025).

Discussion and conclusions

Cross-theme intersections

While the six themes are distinct, several papers sit at intersections, reflecting the reality that deployed digital twins cannot separate architecture, data, security, testing, and application concerns. Doenhoff et al. (2022) bridges security/trust (Theme 4) and modeling/testing infrastructure (Theme 5) by showing that security DTs fail operationally if OT/IoT data collection is fragmented. Cao et al. (2019) bridges modeling/simulation (Theme 5) and applications (Theme 6) by demonstrating that DT credibility for energy systems requires co-evaluating communications, control, and physical dynamics under real-time constraints. Several 6G and edge-network papers bridge frameworks (Theme 5) and domain use cases (Theme 6) by positioning DTs as continuous simulation engines for network planning and orchestration (Khan et al., 2022a; Tang et al., 2022; Karobi et al., 2025).

Novel observations and research agenda

Examining these intersections generates observations that go beyond what existing surveys have identified:

First, the “best” DT design is rarely the highest-fidelity model; it is the model most verifiable under operational constraints. This shows up when DTs connect to control or safety decisions where timing budgets and cyber-physical coupling influence what fidelity is feasible and what validation is mandatory (Cao et al., 2019; Dai and Burns, 2020). Existing surveys that focus on technology catalogs or threat taxonomies miss this fundamental computer engineering trade-off.

Second, the intersection of interoperability (Theme 5) and application scale (Theme 6) surfaces a core research question not addressed by prior surveys: Which interoperability layer actually limits DT transfer across domains: data format, semantics, behavior, timing, or service composition? Cross-domain work suggests domain-specific DTs repeatedly reinvent similar patterns but cannot reliably interoperate due to mismatched assumptions (Acharya et al., 2024; Xiong et al., 2025).

Third, the intersection of trust/security (Theme 4) with adaptive intelligence (Theme 3) highlights a critical tension: DTs increasingly rely on online learning, but that adaptivity introduces drift, stability risks, and verification complexity. This motivates research into safe continual-learning DT pipelines with standardized evaluation protocols measuring detection latency, adaptation stability, and failure modes under distribution shift (Gupta et al., 2024; Shimonishi et al., 2024). Prior surveys identify security and AI as separate concerns; this review reveals their entanglement as a computer engineering design problem.

Fourth, surveys and conceptual frameworks are abundant, but fewer studies provide comparable, testable, reusable artifacts such as co-simulation testbeds, open-source framework baselines, or standardized benchmarks. This imbalance suggests an opportunity to elevate DT research from isolated demonstrations to larger-scale, comparative investigations under consistent criteria (Gil et al., 2024; Somers et al., 2023). While energy systems and Industry 4.0 are strongly represented, human-centered DTs, accessibility-relevant CPS, and broader socio-technical deployments remain comparatively less visible (Rocha-Jacome et al., 2022; Zhang et al., 2022).

Fifth, blockchain and distributed ledger technologies represent a notable emerging but underdeveloped direction for computer-engineering-oriented DT research. These mechanisms are potentially relevant when multiple organizations, devices, or twins must exchange data without a single trusted authority, especially in smart grids, industrial IoT ecosystems, and multi-domain DT networks. At the same time, ledger-based solutions must be evaluated against latency, scalability, privacy, and energy constraints before they can be recommended for safety-critical DT deployments (Hemdan et al., 2023; Somma et al., 2024; Ali et al., 2025).

Synthesis of recommendations

Across the reviewed literature, several high-level guidelines for effective DT development emerge: (1) adopt explicit DT architectures and reference models to support modularity and scalability (Alam and El Saddik, 2017; Fuller et al., 2020); (2) prioritize interoperability as an engineering target, including semantics and timing, not just data formatting (Acharya et al., 2024; Sciullo et al., 2024); (3) treat verification, testing, and co-simulation as first-class DT functions (Somers et al., 2023; Cao et al., 2019); (4) design DTs to be trustworthy under uncertainty, security threats, and evolving conditions (Doenhoff et al., 2022; Shimonishi et al., 2024); (5) integrate intelligent methods carefully, accounting for drift, safety, and deployment constraints as DTs move toward agentic behaviors (Gupta et al., 2024; Antonesi et al., 2025); and (6) evaluate blockchain or distributed-ledger mechanisms when DTs require decentralized trust, provenance, and auditable cross-organization data sharing, while explicitly testing their latency and scalability costs.

Limitations

The final article set was selected using exclusion criteria without formal quality appraisal, so methodologically strong and less robust studies receive equal treatment. Consistent with prior review studies in engineering education and related technical fields (Harris and Kittur, 2024; Kittur et al., 2024; Nguyen and Kittur, 2024), this study limited the literature search to six major databases that were considered most likely to index relevant journal and conference publications. Nevertheless, it is possible that some relevant studies were not captured due to the restriction to these selected databases. Limiting inclusion to articles published from 2015 onward, written in English, and available as open access may have omitted relevant work, including earlier foundational contributions, non-English scholarship, paywalled studies from major publishers such as IEEE, ACM, and Springer, and proprietary industry reports. The open-access criterion improved transparency and reproducibility of full-text coding, but it may also have overrepresented journals and venues with open-access availability. The search strategy relied on specific keyword combinations; alternative terminology may not have been captured. Scope was limited to computer and electrical engineering, and theme assignment involved researcher judgment. This SLR should be viewed as a structured but not exhaustive account of scholarship on digital twin technology in computer engineering.

Conclusion

This review demonstrates that digital twins in computer engineering are converging toward system-level, verifiable, interoperable, and increasingly intelligent CPS infrastructure. The biggest remaining gaps center on standardized validation, cross-domain interoperability, and safe adaptivity at scale. By providing a computer engineering systems lens that complements existing technology-focused, threat-focused, and domain-focused surveys, this review offers researchers and practitioners a structured map of where the field stands and where it needs to go.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this study the authors used ChatGPT to improve the language and readability. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Aravinda S. Rao, RNS Institute of Technology, India

Reviewed by: Babu Pillai, Griffith University, Australia

Asif Ahmed Sardar, Curtin University, Australia