1. Introduction
The concept of a digital twin has gained attention of the industrial and research communities, with its presence in the literature steadily increasing. While the term “digital twin” is credited to Grieves [
1], the underlying concept has been explored and applied under various names for decades, if not centuries. Fuller et al. [
2] provided six different definitions of a digital twin offered by various authors. At the heart of all these definitions lies the integration of a physical system with its corresponding model.
The literature identifies three closely related concepts: digital model, digital shadow, and digital twin, as depicted in
Fig. 1 (Fuller et al. [
2]). A digital model refers to a model developed using physics, mathematics, or statistics, applied to products, processes, or systems. Examples include Newton's second law of motion (
F =
ma;
F is the force,
m is the mass, and
a is the acceleration) and the transportation model used in operations research (
Fig. 1(a)). These models utilize data to generate outcomes, but the data flow—specifically the values of independent variables—remains passive, with no active input to the process.
A digital shadow extends this idea, incorporating process data into the model. While the data feeds the model, its output does not directly affect or control the underlying process, as shown in
Fig. 1(b). A digital twin, on the other hand, represents a more advanced integration, as seen in the concepts of model-based control and digital replica. In this case, the model and the process are tightly coupled, with continuous data input and output in real-time (
Fig. 1(c)). Notably, the frequency of data exchange in a digital twin usually exceeds that of the digital model and the digital shadow, offering a more dynamic and responsive interaction.
A comprehensive review of the definitions and applications of design twins was published by Barricelli et al. [
3]. Future research directions and challenges facing digital twins were discussed. Thelen et al. [
4] classified the modeling technologies into physical-to-virtual and virtual-to-physical. The future trends in the digital twin technology were outlined. Wright and Davidson [
5] addressed the vagueness and inconsistencies across different definitions of a model and a digital twin. Key benefits of digital twins were discussed. The evolutionary perspective of a digital twin was presented in Zhang et al. [
6]. Digital twin platforms and the relationship between digital twin and related technologies were discussed.
The progression from a digital model to a digital twin is defined with seven illustrative characteristics in
Fig. 2.
The satisfaction level of a digital model measured in the characteristics listed in
Fig. 2 is low, as opposed to the digital twin. The coverage by these characteristics is application dependent.
2. The literature on design of digital twins
Human et al. [
7] presented a design framework to facilitate decision-making in digital twin development. Their framework consists of six steps: ① analyzing needs and constraints, ② examining modularity of the physical system, ③ allocating services, ④ assessing performance and quality, ⑤ implementing the design, and ⑥ verifying outcomes. This framework was demonstrated through two applications: a heliostat field and a smart city design.
Recognizing the dependency of a digital twin on specific use cases, Semeraro et al. [
8] proposed an approach featuring invariant modeling patterns. These patterns are autonomously extracted from data and can be reused for similar modeling tasks, with industrial applications illustrating the concept presented.
Bowman et al. [
9] advocated for a unified approach to digital twin design, contrasting methodologies focused on single-product twins. Their research examined case studies of digital twins utilizing a systems-based architecture.
Wang et al. [
10] explored digital twin design from a model-based systems engineering perspective, introducing a framework for estimating design effort and complexity while predicting change propagation.
Redelinghuys et al. [
11] proposed a six-layer digital twin architecture for seamless data exchange, applicable to both legacy and new systems. This architecture was validated through practical implementation.
Moyne et al. [
12] analyzed existing definitions of digital twins and identified key design requirements, including reusability, interoperability, interchangeability, maintainability, extensibility, and autonomy throughout the digital twin lifecycle. Their proposed framework addressed several of these requirements and was demonstrated through case studies.
Lastly, Miller and Spatz [
13] explored the concept of a human digital twin. Their literature review led to a generalized structure for human digital twins, with discussions on applications in product design and sustainment.
Types of models in digital twins: The fact that a model makes the core of a digital twin; the six possible types of the models in manufacturing applications include:
• Machine-learning models
• Statistical models
• Simulation models
• Mathematical programming models
• Physics-based models
• Process models
Model accuracy is key to functioning of digital twins.
3. Development process of a digital twin
The literature explores digital twin applications across a wide range of economic and social domains, including manufacturing, engineering design, healthcare, energy, transportation, and urban development. However, the extent to which these applications align with the true definition of a digital twin varies. A genuine digital twin should leverage data from its operational environment to generate meaningful decisions within that context. Achieving this requires a high degree of model accuracy, although the necessary precision depends on the application—while optimizing city traffic flow may tolerate some approximation, controlling a power plant demands far greater accuracy.
Like any technology, digital twin applications are influenced by factors, such as return on investment, technical feasibility, and data availability. To ensure successful implementation, a structured methodology is necessary. This methodology is likely to share similarities with those used in the design of products, processes, and systems, yet its scope and level of granularity will depend on the specific application and the expertise of the project team.
The following tasks typically constitute a digital twin design methodology. While some tasks may be omitted or modified depending on the project and design team, others may be performed simultaneously or iteratively, rather than in strict sequence.
Fig. 3 illustrates the development process of a digital twin involving 15 tasks that are discussed next.
Task 1: Application feasibility study
Developing a digital twin is a complex and often informal process, necessitating a thorough feasibility study beforehand. A literature review provides valuable insights into past developments within the application domain, including complexity, scope, and success rates. However, specific use cases are rarely documented. The outcome of this task largely depends on access to digital libraries, with search efficiency influenced by the analyst's skills and chosen methodology.
A crucial aspect of Task 1 is cost analysis, particularly the return on investment. However, this analysis cannot be reliably performed before completing portions of Task 2 and Task 3. Several iterations between these tasks are expected, and feedback loops involving other downstream tasks may also arise.
Task 2: Model selection
Selecting an appropriate model for a digital twin depends on the nature of the application and data availability. In some cases, multiple models may be used. Since all models require data, Task 2 and Task 3 are often considered concurrently. Typically, models are built using existing data, which, while practical, may impose limitations on fidelity. To enhance accuracy, efforts should be made to collect additional data whenever feasible.
Task 3: Data engineering
Data engineering encompasses various activities that facilitate data generation to support the selected model in Task 2. The scope of data engineering depends on the model type. For example, machine-learning models require careful parameter selection and quality assessments.
Data engineering is an evolving discipline, shaped by multiple advancements. Mori and Cleve [
14] explored large-volume data processing, while Solanki et al. [
15] advocated for bridging software and data engineering through ontologies that improve interoperability and reduce complexity. Wilde et al. [
16] highlighted the role of ontologies in lifecycle engineering, while Zhou et al. [
17] proposed interaction-based feature selection metrics. Despite these advancements, data engineering still relies on intuition and experience-based methods for selecting parameters, sometimes requiring the installation of new sensors—a factor that may necessitate revisiting Task 1.
Task 4: Model building
Model building involves developing one or more models that accurately represent the physical system. Each model type has distinct properties and requires specific construction methods as outlined next.
(1) Machine learning models are highly accurate and frequently used in manufacturing digital twins (e.g., process control).
(2) Statistical models tend to be less accurate but remain useful for decision-making applications.
(3) Simulation models (discrete or continuous) evaluate different scenarios, such as traffic control systems.
(4) Mathematical programming models assist in optimization tasks, such as material composition for a melting furnace.
(5) Physics-based models are a subset of mathematical programming models but often less complex.
(6) Process models integrate various model types and are well-suited for complex applications.
Task 5: Model validation
Model validation assesses the accuracy of the developed model, typically using the data employed in its creation. Cross-validation is a widely used technique for ensuring reliability.
Task 6: Model testing
Model testing involves evaluating the model using unseen test data. This step identifies gaps that may have been overlooked during validation. A high cross-validation accuracy does not necessarily ensure high test accuracy, especially in non-stationary processes, where data characteristics change over time.
Task 7: Model deployment
Before full implementation, the validated and tested model can be deployed on operational data, allowing for potential adjustments.
Task 8: Model evolution
Not all digital twins require model evolution, but certain models degrade over time. Periodic updates or scheduled model regeneration ensure continued validity, with agentic technologies (Kusiak [
18] in 2025) offering automated approaches.
Task 9: Digital twin specifications
Design requirements for a digital twin should be considered early in the development process. Since no preceding task dictates this step, it may be initiated at any time and revisited as needed.
Task 10: Digital twin design
The model developed in Task 4 and refined through subsequent tasks forms the foundation of the digital twin, which is integrated into its operational environment during this stage.
Task 11: Explainable artificial intelligence (XAI)
Machine learning models, particularly those generated through deep learning, often lack transparency. Users and analysts may struggle to interpret results, necessitating XAI techniques to enhance understanding and model interpretability.
Task 12: Digital twin testing
Although the model has been tested and validated, the digital twin itself must undergo independent testing.
Task 13: Digital twin implementation
Once testing confirms proper functionality, implementation proceeds.
Task 14: Digital twin operations
The digital twin operates as long as it continues to meet its intended requirements.
Task 15: Digital twin maintenance
Routine and need-based maintenance ensures ongoing reliability and performance [
19].
4. Conclusion
Despite the abundance of literature on digital twins, insights into their development process remain limited. This paper captures the digital twin development process in 15 distinct tasks. The proposed model provides valuable insights and highlights the complexity of digital twins. It is designed as a sample roadmap that can be adapted for various applications.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Recommended articles