Model-Driven Engineering of Digital Twins

There are two core terms that need appropriate definitions, namely model driven engineering (MDE) and digital twin. While MDE (and with it the terms model and modeling language) are relatively straightforward, there are different variants of definitions for digital twins that served as a starting base.

As digital twins are currently a relatively young and in particular evolving area of research, it is not surprising, that there is a variety of different definitions available. (E.g. [1] identifies more than hundred different definitions).

At the beginning and during the seminar we identified the following definitions to be particularly of interest.

An alternative definition can be found at the website of the software engineering institute of RWTH Aachen University⁴⁴4https://www.se-rwth.de/essay/Digital-Twin-Definition/ together with some additional discussions:

Figure 1 depicts the schedule of the seminar week. It started by opening the seminar week and then with presentations and discussions about the terminology, context, and properties of digital twins. We then discussed the planning of the rest of the week. In the second day, we had presentations of real world scenarios of digital twins. After that, we discussed about requirements for digital twins and finalized the planning of the rest of the week. Finally, for the rest of the week, we split in three groups to work in parallel, with a plenary session in the morning and the evening to debrief each other about what was done and will be done.

Given that operational cyber-physical systems (CPS) produce continuous data, a complementary approach to model-based engineering is to learn digital twins models with machine learning techniques and providing functionalities such as predictions and anomaly detection.

This talk will start with presenting an opinion on the next generation of digital twins (Quantum Digital Twins), where some aspects of digital twins will be implemented as quantum software and executed on quantum computers, e.g., for simulating the physical environment that can be realistically simulated with quantum-mechanical principles.

Followed by this opinion, the talk will present some recent works on learning digital twins from historical data and continuous updates of digital twins with continuous data from operational CPS. Various machine learning techniques were applied, such as generative adversarial networks, curriculum learning, and transfer learning to learn digital twins. The digital twins were built for use cases from the transportation domain and water distribution/treatment plants. These digital twins were focused on anomaly detection and waiting time predictions.

Depending upon the flavour of digital twin (and that is an entirely different problem), a digital twin of a real world artefact is designed for a purpose. Such a purpose can can range from trying to understand an “as is” situation, analysing multiple options in a decision making scenario through to tractable transformation of the real world artefact. Underpinning all these purposes is a form of theory building and assumptions of underlying theories.

A theory in its most sparse understanding is a statement of what causes what, and why, and under what circumstances. A theory can be a contingent statement or a proven statement. We use theories all the time. Peter Naur referred to programming as form of theory construction. Decision makers in organisations use theory every day. They make decisions on some basis of cause and effect, often without being specific about their reasoning. Naturally, theories are empirical or can remain conceptual explanations.

The most widely accepted notion of a Digital Twin (DT) is one where there is a cyber-physical component. DTs of civil structures such as bridges are one such example. Such DTs rely on well understood physics based laws which provide empirically tested theories. New generations of DT belong to a socio-technical context where there is heady mix of human action, systems integration and emergent behaviour that cannot readily be assumed to follow a particular predicted route. Such Socio-technical DTs when they are constructed need input from social science, psychology and other inter-disciplinary fields. Theories for justifying concepts, relationships and rules of such a STDT are therefore obtained from these non-IT fields. How do we embed and use these theories?

The challenge for model driven engineering of STDT is therefore further attenuated and more engineering support is needed. For example, we need a modelling language that is sufficiently rich to capture working descriptions of key social science theories. We will need libraries of these theories together with working examples. We will need theory integration environments and accompanying methods that support theory building. We want to be able to run a model and view the execution of the theory and its explanation.

The benefits of being able to demonstrably prove that an established theory underpins a component of a DT specification model provides external validity through a reference bench mark of a well understood theory that has been critiqued at scale and over time. Model validity concerns are therefore mitigated. Essentially this is a research challenge and I am not yet aware of significant progress in this area with respect to Digital Twins.

Advancements in the automotive industry is being highly dependent on Software technology as the industry is now changing from manual driving to different levels of autonomous driving. In order to develop autonomous driving vehicles, digital twins are now used to achieve the desired features and functionality. The approach to use digital twins saves huge resources as it accelerates product development and perform virtual simulation forecasts before production. Another use case is using the twin in tandem with the physical environment providing functionalities of monitoring, fault detection to name a few. The Automotive Technology research lab at Eindhoven University of Technology, Netherlands is performing research and development of such systems. As a part of this lab, this project focuses on design and implementation of a digital twin software framework for autonomous articulated vehicles within a distribution center. The goal is to develop such a system, that replicates ‘the real world moving of autonomous trucks in a distribution center’ within a Digital Twin. Additionally, the virtual system should be able to gather data from various sensors in order to enhance the development of the physical truck. This data can then be used by users to further study or visualize. In order to develop this complex system, a software system representing the digital twin called the TruckLab-DTF has been developed. Using the TruckLab-DTF product development teams working on achieving autonomy in distribution centers can apply insights from the virtual twin and physical system directly to their development. Additionally, requirements can be verified in the digital twin early in the design phase, saving time and money. Thus, the use of digital twin with this domain is multi and this project is an initial step towards the realization of much larger “system of systems.” In extension, designed software can be also used as an educational technology tool for vehicle dynamics and control-based courses.

Digital twins (DT) have emerged as a promising paradigm for run-time modelling and performability prediction of cyber-physical systems (CPS) that can be applied in multiple domains. Different definitions and industrial applications of DT have materialised, going from purely visual three-dimensional models to predictive tools, many of them focusing on data-driven evaluations. We want to focus on a conceptual framework based on autonomic systems to host DT run-time models based on a structured and systematic approach. A model at run-time can be defined as an abstract representation of a system, including its structure and behaviour, which exist alongside the running system. Run-time models provide support for decision-making and reasoning based on design-time knowledge. However, they can also offer themselves as a run-time abstraction based on information that may emerge at run-time and was not foreseen before execution. New techniques based on machine learning (ML) and Bayesian inference offer great potential to support the update of run-time models during execution. Run-time models can be updated using these new techniques to provide better-informed decision-making based on evidence collected at run-time. The syncing of the real and the digital twin: Models@runt.time and the MAPE-K loop can provide the structured basis of the software architecture presented and how the required casual connection of run-time models is realised to sync the real and the digital twins. This process keeps the running system inside an envelope of good behaviour.

DevOps emerged in the last decade as the prominent approach to increase productivity and system quality in the software industry. It advocates for automation and monitoring at all stages of software development and operations, and aims for shorter development cycles, increased frequency of deployment, and more reliable releases. Its adoption by industry leaders (e.g. Amazon, Facebook, Google, and Netflix) has resulted in spectacular progress. However, evolving/improving the software process remains a main challenge and many companies are struggling with the implementation and evolution of software processes. The lack of a systematic approach makes continuous improvement an ad hoc journey in which decisions are based on intuition rather than facts. To enable the systematic improvement of DevOps processes, we propose a digital twin approach that addresses two main challenges: 1) the continuous monitoring and measurement of DevOps process according to specific objectives to detect issues so that improvement actions can be taken; and 2) the evaluation of various modification alternatives to reach a specified DevOps process improvement objective; this allows to take decisions on a scientific basis rather than in an ad hoc manner.

A series of architecture models for digital twins of increasing sophistication is presented, leading to a digital twin that adapts in order to control a real-world asset. A technology is in development that supports the construction of adaptable digital twins. An overview of a simple use of the technology is presented.

Enterprises are rapidly evolving as complex socio-technical systems that require continuous adaptation regarding a dynamic and uncertain environment. Various stakeholders are involved and deliver continuous knowledge through heterogeneous digital models (systems, processes, organizations, etc.) manipulated with various scientific and engineering environments. This modeling continuum open new opportunities for adaptable and efficient enterprises, but also raise new modeling challenges. In this talk, I explore the use of model-driven engineering to develop and operate enterprise digital twins. The talk covers the current state of the art, and provide concrete implementations for smart trade-off analysis and decision making. Finally, I discuss open challenges and draw a roadmap for the community.

The talk will give a very brief summary of existing Digital Twin efforts in the US, and then introduces some ideas about the specific challenges digital twins have to solve in the domain of Cyber-Physical Systems. A specific approach for addressing the model integration problem is discussed, as well as how DT-s can assist in autonomous system operations. Finally, five fundamental challenges are posed.

Product complexity is ubiquitous – and it is constantly increasing. Instead of working against it, we should accept it as challenge and make complexity manageable.

Why are software methods becoming increasingly important in this context? Because that is exactly where they come in: Complexity becomes controllable by software methods. For the automotive business, this means that we must use systems engineering methods to manage product variance holistically, based on a central data model having an end-to-end scope over lifetime. To do this, we follow a data-centric approach that maps every aspect: Both in the problem space (requirements, features, regulatory specifications) including the product configurations and in the solution space (types, options, functions, components) including the corresponding type configurations.

In this talk an overview of our concept “Typebased Product Line Engineering” is presented: How can we move from a reactive, quantity-based product documentation approach to a proactive product documentation approach based on individual and concrete configurations? The role of a Digital Twin is outlined – especially in the context of a vehicle that can always be upgraded with new features via over-the-air updates throughout its lifetime in customer hands.

Acknowledgement: The talk was originally presented by Christian Seiler, Mercedes-Benz AG at the Digital Product Forum 2022. This work is partially based on the research project SofDCar (19S21002), which is funded by the German Federal Ministry for Economic Affairs and Climate Action.

Deployment denotes the installation and the running of necessary components of an application. There are mulitple tools available to do so: Terraform nad Ansible form two examples. They all have different meta-models to capture the desired application topology (also called “infrastrucutre-as-code”). OASIS TOSCA (“Topology and Orchestration Specification for Cloud Applications”) is a standard including both a meta-model and a meta-meta model to model the application topology in a standardized and vendor-neutral way. This talk outlines aspects of TOSCA and presents Eclipse Winery as one tool to model applicaton topologies using TOSCA.

Acknowledgement: This work is partially based on the research project SofDCar (19S21002), which is funded by the German Federal Ministry for Economic Affairs and Climate Action.

JabRef is a literature management software based on BibTeX written in Java. The project management is mainly done using GitHub’s features. The talk will fist outlined JabRef’s functionalities and then dives into open source software development. We showed how the JabRef team supports newcomers to find new issues and to craft a code contribution – and how this helps to use JabRef as teaching object for training Software Engineering on intermediate level.

With the emergence of Digital Twins, more and more architectures are developed to use these Digital Twins in particular context to serve a specific purpose. However, the components of such architectures are usually targeted to its particular context and purpose, with limited adaptability. This makes it hard to change an architecture once the purpose of the underlying system changes, or reuse existing components in developing new architectures. To solve this challenge, we propose a feature model of Digital Twins that enables to put together new software architectures that leverage these Digital Twins. The component-based realization of the individual features on the software and modeling language level should enable an efficient plug and use of Digital Twin architectures in the future.

From a data-centric point of view digital twins are just data structures/models that keep relevant information of real-world artifacts (machines, products, environments etc.). Thus “working with digital twins (cf.my title)” simply means adapting/changing/extending/merging/intersecting/snipping/… these structures/models. My main questions raised – and perhaps even answered during this seminar – are:

• $\mathrm{■}$

  Is this really so simple?

• $\mathrm{■}$

  What does it mean to merge two digital twin models and what is the impact and consequences thereof?

• $\mathrm{■}$

  Don’t we need to consider (domain) semantics?

The concept of classification as realized in most traditional object-oriented computer languages has certain limitations that may inhibit its application to modeling more complex phenomena. This is likely to prove problematic as modern software becomes increasingly more integrated with the highly dynamic physical world. In this paper, we first provide a detailed description of these limitations, followed by an outline of a novel approach to classification designed to overcome them. The proposed approach replaces the static multiple-inheritance hierarchy approach found in many object-oriented languages with multiple dynamic class hierarchies each based on different classification criteria. Furthermore, to better deal with ambiguous classification schemes, it supports potentially overlapping class membership within any given scheme. Also included is a brief overview of how this approach could be realized in the design of advanced computer languages.

We have already explored engineering a simulation so that it can be trusted (in the non-strict sense), and so that the basis for the trust can be interrogated and challenged: the principled simulation approach has been used in e.g. immune system simulation and robotics. In 2000s we proposed MDE as one way to automate development so that effort could focus on abstraction, design and interpretation of results, plus use of the simulation for its purpose. Digital Twin development might perhaps learn from engineering principled simulation, particularly in areas such as purpose, results interpretation/use, fitness/validation, management of trustworthiness. Again MDE could allow “people” to focus on these non-automatable but critical areas.

This talk describes challenges for the development of the digital twin from a modelling perspective.

Manufacturing systems are structured as multi-layered systems, with different goals and constraints for the layers. Digital Twins must reflect this by being split into parts what we call Federated Digital Twin. For its parts, and for the bridges between them, appropriate decisions on coverage, structure, technology, etc. are required.

Digital Twins need to accommodate the interactions of the System under Study (SuS) to be able to collect data, stimuli and outputs. This requires interaction with the (software) ecosystem surrounding the SuS, determined by relevant functions, and defining a landscape of models and their interactions.

A fundamental characteristic of software models is their ability to represent the relevant characteristics of the system under study, at the appropriate level of abstraction. We now live in the age of cyber-physical systems, smart applications and the Internet of things, which require some forms of interaction with the physical world. Uncertainty is an inherent property of any system that operates in a real environment or that interacts with physical elements or with humans but, unfortunately, the explicit representation, management and analysis of uncertainty has not received much attention by the software modeling community. In this talk we analyze the impact of measurement uncertainty on the behavior of the system using examples, and describe the traditional ways of dealing with it, namely using fixed and confidence (adaptive) intervals. We then discuss the need to consider all attributes as random variables and the importance of properly comparing them.

Throughout their life-cycle (design, production, assembly, operation and optimization, maintenance, re-purposing, disposal), smart, adaptive systems will include ecosystems of Digital Zs (models/shadows/twins) to achieve a plethora of goals such as condition monitoring, fault diagnosis, predictive maintenance, optimization, etc. . This builds on many existing techniques, architectures and standards from real-time simulation, co-simulation, systems and control theory, IoT, knowledge management, machine learning, experiment/validity frames, surrogate modelling, etc. To satisfy system goals, a federated knowledge repository (graph) (a “Modelverse”) containing both “linguistic” and “ontological” information, is used as a starting point for inferencing. This leads to the (product line: goals to realizations) construction of new Digital Z “experiments” which, when deployed, in turn, yield new data/knowledge which is merged (during or at the end of an experiment) with the information already present in the Modelverse. As multiple concurrent inferencing and experiment processes may exist, the Modelverse acts as a blackboard. Conceptually, one Digital Z “experiment” is created per Property of Interest (to be monitored, satisfied or optimized).

Multi-Paradigm Modelling (MPM) principles are used throughout: model explicitly, most appropriate formalism(s), abstraction(s) for architectures and views, requirements and designs, with workflows modelled explicitly too.

Digital twins are starting to appear everywhere. In this talk, I reflect upon definitions, purposes, and tools related to engineering and operating digital twins based on the largest cross-domain mapping study on the topic to date.

I report on our experience building agent-based simulations and, further, digital twins for learning healthcare systems, the role that MDE has to play in this, and the challenges for future work.

Digital Twins play an important role in engineering complex (high-tech) systems. They allow for real-time analysis of engineered systems in order to detect anomalies, to predict maintenance, and to optimise behaviour. The use of model-driven techniques accelerates the development of digital twins. The focus of the Software Engineering and Technology group from the TU/e focuses on the efficient development of Digital Twins, by means of model reuse and advance orchestration of the reused models.

The use of models used in the engineering process leads to the question whether the digital twin can replace the supervisory control of the engineered system. What are the requirements to facilitate this:

• $\mathrm{■}$

  Supervisory control in the cloud by means of the engineered system

• $\mathrm{■}$

  Fast communication channels between physical entity and the engineered system (5G and beyond)

• $\mathrm{■}$

  Sensors, actuators and edge computation on the engineered system.

In this group, the participants focused on identifying the different contexts in which a Digital Twin may be produced, the goals and purposes of those Digital Twins, and their qualities and properties. We motivated our definitions based on the range of talks given during the seminar, as outlined in section 4. The talks included a variety of contexts for digital twins, including a variety of assets for which a digital twin is constructed, a variety of environments in which the digital twin is embedded, and a variety of purposes for which the digital twin was constructed. Below, we provide a short summary of the key discussion points within the group, and our findings:

From the range of talks demonstrated in section 4, we have seen that there are many different assets for which a digital twin can be developed. We came to define the ‘context’ of a digital twin as the asset type, as well as the asset’s environment. Properties of the environment may include whether there is human intervention in the environment, whether the system is cyber-controlled or not, and whether the asset was engineered or is natural.

In the problem space group, we then discussed the “qualities” of digital twins. I.e. the desirable properties of the system towards some measure or mitigation. We discussed the qualities of both the digital twin as an artifact, as well as the qualities that the digital twin must achieve for the user. Qualities apply to all engineered aspects including I/O and data flows between the physical and digital twins as well as the digital twin itself. We do not, however, comment on the qualities of the ‘physical twin’ as this is beyond the scope of the concerns of a digital twin developer.

During our discussions, we reached the opinion that qualities in the digital twin context are closely linked to qualities in the systems/software engineering context more generally. We, therefore, did not endeavor to produce a comprehensive list of qualities but instead aimed to identify some of the qualities of relevance to (different) digital twins.

In figure 2 we annotate a digital twin architecture diagram with the most relevant quality properties. These include:

• $\mathrm{■}$

  Validity: A measure to show that the digital twin is a “good enough” digital representation of the physical entity/twin

• $\mathrm{■}$

  Fidelity: How much reliance can be placed on the outputs (data, control signals, observations, etc.) of the digital twin.

• $\mathrm{■}$

  Trustworthiness: Closely linked to properties of fidelity and traceability. The degree to which stakeholders trust the validity of the digital twin as a whole.

• $\mathrm{■}$

  Modularity: The capacity to allow for the substitution of components such as services to fit the needs of the user in a given context

• $\mathrm{■}$

  Usability: The human-computer interaction properties of the system, and the standard of user experience the system provide

• $\mathrm{■}$

  Synchronisation: The properties of ‘lag’ between the digital and physical twin

• $\mathrm{■}$

  Security: Including considerations on risk analysis, data protection, and data biases

The goal of this group’s discussions was to identify the fundamental scope and properties of digital twins. These properties were then used to motivate the requirements for the design space group as discussed in Section 6.