Model-driven engineering for digital twins: a systematic mapping study

Digital twins (DTs) are increasingly used to enable applications such as anomaly detection [74], virtual experimentation [4], or reactive planning [S22] for cyber-physical systems (CPSs) [3] in various domains, e.g., manufacturing [67], agriculture [58], or energy management [75] to mention just a few. However, as DTs comprise of multiple and diverse software components, which highly vary based on specific use cases, they are considered as complex software systems [19]. Thus, designing, developing, and operating DTs require considerable development effort [39] as well as communication and cooperation between software experts and domain experts.

Model-driven engineering (MDE) [11] is a prominent approach to handle complexity and facilitate communication by leveraging the abstraction power of models. These models are designed to be human-understandable symbolic models that simultaneously provide machine-readable and processable representations of real-world entities. By applying MDE, these models can be (i) transformed to other models (e.g., Business Process Model and Notation (BPMN) models to Digital Twin Definition Language (DTDL) models [S60]), (ii) used as a basis for code generation (e.g., BPMN models are translated to Java code [S60]), or (iii) interpreted by software systems, e.g., a simulation controller model is directly interpreted to read simulation results and send commands from/to a running Unity simulation [S53]). More precisely, we use the following characterization of MDE throughout this paper.

Bild vergrößern

Several researchers have already discussed the potential benefits of using MDE in the context of DTs [10, 18], and recent MDE applications already provide evidence of the expected benefits [36]. Since DTs are complex software systems, as already mentioned before, the application of MDE in this context may also be referred to as model-driven software engineering (MDSE) [11]. As MDE is the general term that also includes MDSE, we decided to use the term MDE in the remainder of this paper.

To further advance the current state of the art, the research community requires an overview and classification of existing approaches to (i) understand which application contexts of automation techniques yield most promising results that can be built upon, (ii) how different characteristics of DTs influence this application context, and (iii) which application domains currently see the highest level of adoption of MDE for DTs, and which domains are lacking such activities. We expect that such a documentation of the current body of knowledge can be helpful for the adoption of MDE in practical settings for different DT types and domains, and align future research efforts in this area. Although the application of MDE to DTs has been extensively investigated in recent years, such an overview and comparison for general orientation is still missing.

In order to tackle this shortcoming, in this paper, we shed light on these aspects by performing a systematic mapping study [53] on existing applications of MDE automation techniques to DTs. In particular, we performed a keyword-based search in multiple literature databases. Based on the results obtained from the included papers, we classify applications of MDE for DTs and derive challenges that can be tackled by the research community in the future by proposing concrete action points.

The remainder of this paper is structured as follows.

Section 2 gives an overview of related work and further motivates this study. Subsequently, Sect. 3 gives a detailed description of the applied research method, i.e., systematic mapping study. Section 4 elaborates on the results obtained from using this method, and Sect. 5 discusses these results in the context of the stated research questions. Section 6 presents threats to validity of the obtained results. Section 7 identifies the challenges for advancing the application of MDE to DTs and proposes action points to address them. Finally, Sect. 8 concludes the paper.

As described in the previous section, our research interest focuses on identifying and characterizing existing MDE applications to DTs. Given the diverse landscape of DTs [73, 66], it is currently challenging for researchers and practitioners to achieve a comprehensive overview of existing MDE techniques and accompanying modeling artifacts. This makes it essential to investigate not only what has been done for single applications, but also how these applications differ for the diverse DT types and domains.

In the following, we present existing investigations of related aspects, i.e., applications of MDE to different fields (potentially outside the area of DTs), the usage of different DT types, and widespread application domains of DTs. We also highlight gaps in these existing surveys to motivate the need for the survey presented in this paper and summarize the found differences of existing surveys in Table 1.

In the following, we describe the method applied to obtain the results of our mapping study, following established guidelines [53, 37]. The process for conducting our study is shown as an activity diagram in Fig. 1.

This section presents our findings on the bibliometric key facts of the included papers, automation techniques used in the context of DTs in these papers, the DT types to which MDE is applied, and the domains where MDE is applied for engineering DTs. We answer the respective research question at the end of each subsection, summarizing our findings.

For identifying the threats to the validity of our study, we follow the four basic types of validity threats according to Wohlin et al. [72]. In this respect, our study is subject to threats to construct validity, internal validity, and conclusion validity. With respect to the threats to external validity (generalizability), it is important to note that the results of this study can neither be generalized to problem domains other than DTs nor to solution domains other than MDE. As we only considered publications from peer-reviewed venues in our study, the results cannot be generalized to applications of MDE or applications to DTs outside academia (i.e., in an industrial context).

Concerning construct validity, our presented findings are only valid for our sample of papers. Thus, we tried to include as many papers as possible from ACM, IEEE, Scopus, and Web of Science, and only excluded papers according to our exclusion criteria. Additionally, we intentionally designed our search query to be relatively broad and inclusive. This approach allowed us to identify papers related to the concept of model-driven approaches without imposing strict limitations, such as exclusively focusing on software engineering or specific MDE techniques. However, it is worth noting potential limitations in our search strategy. Our corpus contains only papers published from 2018 and following. This may be attributed to our search query that included the term “digital twin,” i.e., although researchers had explored analogous concepts before 2018 they did not explicitly label them as digital twins. Furthermore, our search query did not encompass the term “model-based”, and we conducted searches within title, abstract, keywords, and venue title fields, without conducting full-text searches. We also found that the results of the Web of Science database seem to depend on the subscription of a research institution to the Web of Science core database, as also described by other researchers [31, 56]. To be as inclusive as possible, we used all search results obtained from the subscriptions of two different universities, namely from the Johannes Kepler University Linz and the University of Stuttgart. To enhance the comprehensiveness of our paper selection, we conducted forward and backward snowballing, aiming to identify papers that might not have been captured by our initial search query. Another potential threat to the construct validity of our study pertains to our chosen inclusion and exclusion criteria, which were applied solely to the title, abstract, and keywords during the screening phase. To minimize the risk of inadvertently excluding relevant publications, we included papers where uncertainty existed among one or more authors, subjecting them to further investigation during the detailed review process. Throughout both the screening and detailed review phases, a collaborative effort involving multiple authors was employed to read and discuss the papers, ensuring a rigorous evaluation process. Excluding philosophical, opinion, and experience papers in this process has resulted in 10 exclusions during the screening and 20 exclusions during the detailed reviewing phase. Even though these papers could have included more diverse perspectives in our results, we decided to only include publications that provide a significant level of technical contribution and description which can be analyzed to extract the necessary information for this study.

Our mapping study is also subject to the so-called publication bias [29], i.e., our study can only report on published results. This limitation implies that our study may not encompass all relevant research, as unpublished or non-peer-reviewed work is not included. For example, since our focus was on peer-reviewed literature, we did not find industrial solutions, such as the Azure Digital Twins service,⁵ or the Ansys Twin Builder,⁶ which might have been accessible via blog posts or preprint articles. Including such publications in a multi-vocal literature review [27] might be one line of future work, as this gray literature could include solutions with higher technological readiness levels, such as the examples given above. Additionally, we included publications that claim to utilize MDE without providing extensive examples. In such cases, we relied on the authors’ integrity and adherence to scientific principles, as we must trust the accuracy of their claims. Another aspect of construct validity is the so-called mono-method bias, referring to the use of a single metric to analyze the selected papers. We counteracted this bias by providing several metrics to answer each of our research questions. To mitigate the issue that one paper might apply several aspects that are relevant for our study, i.e., several automation techniques, we extracted all automation techniques for each paper as a dedicated entry. To cope with the option that an extracted automation technique regarded aspects of several categories for a particular metric, we allowed the assignment of automation techniques to multiple categories for each metric.

A threat to internal validity is that our study relies on the quality of papers that were included. Additionally, our categorization may not comprehensively encompass all facets of the described data sets. This limitation stems from the fact that certain papers offer only high-level overviews, while others provide extensive details.

Threats to conclusion validity arise drawing wrong conclusions and from the study’s replicability. Regarding the former, we have discussed issues that could lead to wrong conclusions in the context of threats to internal validity. For study replicability, we describe the method we used to derive the described results in Sect. 3 and provide a replication package containing all primary and secondary data, as well as the automation scripts we used to process this data, online.⁷

In the following, we present a list of challenges which have to be tackled in order to further advance the application of MDE to DTs based on the results of our survey. In addition, we propose several action points to tackle these challenges in future activities.

In areas that already contain a high number of evaluation research (e.g., applications of model interpretation (RQ1.4), or applications to DTs representing individual systems (RQ2.2)), it seems promising to build upon this existing work by replicating evaluations for different application contexts, DT kinds, and domains, or reuse the tools described in these papers for future work. For the manufacturing and transport domain, there are also many real-life use cases available that researchers can build upon, as described in detail in RQ3.1. However, even though data from these use cases are available, applying automation techniques to the physical systems used in these domains (e.g., injection molding machines, transportation systems) might be rather cost-intensive and require field access by the researchers. Thus, the construction domain also seems to be a good target for reusing existing solutions, as most presented use cases use low-cost equipment (floor cleaning robots or air quality sensors) that can easily be rebuilt in lab environments. To share access to high-cost physical systems, we need cross-disciplinary research between communities that already perform research in domains such as manufacturing or transport and communities that can build DTs and automation techniques for available systems.

Action Points. The adoption of existing research highly depends on the usage of common languages. As DTs combine different models from different categories, such as behavioral models used by simulations, data models used by data processing or storage solutions, or UI models used by graphical user interfaces, we need a family of DT languages covering these different concerns, that can also be flexibly adapted based on the DT kind (cf. RQ 2) [55]. Common languages for these different concerns are currently being defined by consortia built from both industrial and academic members, such as the Industrial Digital Twin Association⁸ or the Digital Twin Consortium⁹ by (i) extending existing general-purpose modeling languages such as UML, SysML, and BPMN, (ii) adapting existing domain-specific languages for DTs, such as the Asset Administration Shell, to the specific requirements of MDE for DTs research, or (iii) building new languages from scratch. An investigation of existing languages of DT platforms showed that extending languages that are already used to engineer twinned systems, such as UML class diagrams, provides a solid basis for such common languages [54]. One advantage of reusing such languages is that we can also reuse existing models defined in these languages that already contain information on the twinned system for generating the DT of this system [S38], or reuse existing MDE tooling that is built based on these languages.

Challenge: Strengthen the connection between academic and industrial DT applications. We also note that our study explicitly focused on peer-reviewed publications. Some of the identified clusters of MDE applications might be different in a purely industrial context or applications that are not published in peer-reviewed venues. For example, we observe in our research projects with industrial partners that several industrial DT applications currently make use of languages such as the Asset Administration Shell or Digital Twin Definition Language, and there are also emerging interpretation engines available for these languages, such as the Azure digital twins explorer¹⁰ or Eclipse BaSyx project.¹¹ We also observe a drift between the above-mentioned technology, and the tools and languages used in academia.

Action Points. Investigating applications of MDE in non-peer-reviewed literature can be used to confirm, contradict, or even extend the results presented in this paper. Comparing industrial DT practices to the scientific results presented in this study can show interesting gaps between research and practice. To survey non-peer-reviewed literature, a follow-up study can, e.g., make use of the guidelines proposed by Garousi et al. [27]. Additionally, researchers can perform interview studies with industry and domain experts to understand what these stakeholders need, and what is currently missing in the state of the art to fulfill those needs. To better understand industrial requirements, we can also investigate further features and languages used by existing DT technologies by extending the work by Qi et al [59] and Lehner et al. [40]. However, as these technologies are constantly evolving, it also makes sense to provide dedicated online repositories or websites to monitor these platforms.¹² Besides industrial applications, we can also explore DT-specific standards, such as the Asset Administration Shell or the ISO 23247 Digital Twin Framework for Manufacturing [32], and compare them with existing standards, such as UML, SysML, and AutomationML. On a technological side, we can provide bridges between these different modeling languages proposed by these standards to allow transitioning between them.

Action Points. We uncovered that a considerable amount of automation techniques is transforming data or discrete behavior models into a different representation of the same category, e.g., for refinement or interpretation of behavior models using simulators for experimental investigations. Using different combinations of source and target artifacts, or broadening the applications of certain model categories, might lead to interesting new insights. For example, enriching DTs with semantic information seems promising [76, 9], but MDE is currently applied to ontologies for rather limited purposes, i.e., mostly to derive data models or configure control or simulation software. For model interpretation and code generation in particular, using different model types increases the information that is available to the software used as generation target or interpretation engine, and thus extends the possible range of applications. For example, behavioral models are currently rarely used for code generation, in addition to generating simulation code, e.g., to automatically generate integration code that connects different components of DT architectures, or to allow interaction with the dynamic behavior of the PT using visualization or control software.

Action Points. To tackle this challenge, we can make use of existing proposals, such as applying DTs to processes [13] to provide solutions proposals that can then be further validated and evaluated. We can also reuse existing DTs, and start applying automation techniques to them. For example, DTs have already been applied to biological beings, such as humans [6] or cows [50], or overlapping categories such as socio-technical DTs [5] and cyber-biophysical systems [20]. With respect to the SLCP of the twinned system, the research community has already gathered a considerable amount of knowledge on how automation techniques are applied in the testing [69] or maintenance [70] phases of a system. Another aspect which seems not to be investigated until now is the application of MDE for the requirements phases of DTs.

Challenge: Adopt MDE for DTs in new domains benefiting from digitalization. The results of RQ3 show a large imbalance with respect to the domains to which MDE is currently applied. More than half of the papers focus on two domains only. Thus, the generalization of MDE approaches applied in these domains seems to be a critical research direction for the MDE community as modeling is considered as a universal discipline [15]. In manufacturing, modeling languages are already widely used [73], and Industry 4.0 (having the RAMI reference framework) and Industry 5.0 (mentioning DTs as one base technology) seem to be additional initiatives that drive the adoption of DTs and applications of automation techniques to DTs. In other domains, the level of digitalization to realize DT applications might still be lower. In general, when adopting MDE in new domains, we might be faced with technological boundaries, i.e., connecting existing MDE tools to domain-specific standards and modeling languages, as these domains often still lack basic technologies, such as widely adopted middleware and modeling languages, which are, e.g., available in manufacturing with OPCUA and AutomationML.

Action Points. For many domains (that is, 10 out of the 21 investigated domains), the application of MDE has not yet been explored in peer-reviewed publications. For these domains, generalizing existing approaches from more prominent domains seems promising. For domains that already contain a certain amount of MDE applications, such as construction, there are still some unexplored areas with respect to the application context of automation techniques. For example, there is only one application of model-to-model transformation to the transport domain. Automation techniques in the transport domain also currently focus on the operation phase, and applications in the construction domain focus on implementation. Transferring applications in these life cycle phases to other domains (e.g., implementation automation techniques from construction to transport) can help the research community efficiently close existing gaps. For performing this knowledge transfer, we can collaborate with domain experts, and build bridges between domain-specific tools and existing modeling languages and software artifacts used for applying automation techniques to DTs. Interdisciplinary research communities, such as the Engineering Digital Twins community,¹³ can be a good starting point for such initiatives. However, we can also leverage existing surveys [35, 28] or vision papers [64, 63] on potentials for DTs in certain domains as a starting point to apply automation techniques to these DTs.

Challenge: Prevent reinventing the wheel. Modeling languages and the application of automation techniques based on models defined in these languages have been well researched in more than two decades. Even though our survey shows that researchers started implementing automation techniques for DTs in 2018, there is a profound source of knowledge from the MDE community [14], investigating (i) concepts that can be reused for DTs, such as runtime models or (ii) applications of MDE techniques to particular domains in which DTs are also applied. If we disregard this existing knowledge, we run into the risk of reinventing the wheel.

Action Points. Researchers can leverage existing knowledge by investigating existing surveys that describe the application of MDE in related fields, such as Industry 4.0 [73], robotics [16, 1], cyber-physical systems [49], and runtime models [65, 23] as a starting point. Particular action points that researchers should focus on are, e.g., reusing languages and techniques from the runtime models community to perform self-adaptation and connect automation techniques with AI-based methods. From the cyber-physical systems domain, connection of discrete and continuous simulations should be more leveraged for DTs. Researchers may also experiment with applying automation techniques to new sorts of languages, that, e.g., provide 3D modeling capabilities for geometry or kinematics, similar to the COLLADA language which is included in the AutomationML language family.

The systematic mapping study performed in this paper investigated where and how MDE automation techniques have been already applied for DTs. The presented maps illustrate already available clusters of MDE approaches for DTs, but also point out some spots which are interesting areas to work on by researchers from the MDE community, potentially in interdisciplinary settings to target new application domains.

Future research endeavors may focus on building upon the existing solutions that were identified in our study, increase their maturity, and explore their application to other domains, DT kinds, and application contexts. Furthermore, since we identified a variety of employed modeling languages, another research direction is to work toward a unified modeling language or family of languages for DTs. The corpus of techniques and artifacts collected in this study should be a good starting point for this direction. Finally, our study may be extended by including gray literature and industrial applications as well as DT platforms to derive further strategies for enhancing the adoption of MDE for DTs in industry.