Policy Significance Statement

Digital twins (DTs) are an emerging data-centric technology that creates virtual replicas of real-world objects, systems, and individuals. DTs are usually a hybrid of physics-based, process-based, and machine learning models validated by data. As the use of DTs becomes more widespread, it is crucial to establish a new data governance framework to ensure that ethical design issues are addressed. This article advocates for a balanced approach through a mindset of aggregation to scale DTs.

1. A panoply of digital twins

Digital twins (DTs) and their terminology, tools, and techniques have been gathering pace. We collect some recognised definitions to comment on how the concept has evolved through time. Our goal is to emphasise, rather than redefine, emerging themes and summarise them into common traits (Section 2) throughout a hierarchy (Section 3) to discuss a panoply of core technical ideas.

(2002) DTs Introduced: Michael Grieves proposed the DT conceptually, under a different name: a Conceptual Ideal for Product Life Cycle Management (Grieves and Vickers, Reference Grieves, Vickers, Kahlen, Flumerfelt and Alves2017). While general, the concept refers to the three elements that underpin most DT definitions: (i) the real space, (ii) the virtual space, and (iii) the links for information flow between the two. Importantly, the virtual representation is not static; it evolves throughout the system’s life cycle. Considering assets or products, life refers to production (design and manufacture), operation (sustain/support), and disposal. Many Industry 4.0 applications fall under this category of DTs (Kappel et al., Reference Kappel, Brecher, Brockmann and Koren2022).

(2012) Another Core Definition: NASA’s definition (Glaessgen and Stargel, Reference Glaessgen and Stargel2012) is more specific: a DT is an integrated multi-physics, multi-scale, probabilistic simulation of an engineered system—to mirror the life of its physical twin. Critically, the terminology considers interdependent sub-systems, including propulsion/energy storage, avionics, life support, vehicle structure, and thermal management (Glaessgen and Stargel, Reference Glaessgen and Stargel2012; Negri et al., Reference Negri, Fumagalli and Macchi2017). While still focussing on manufactured components, there is consideration of interactions between aggregated systems and virtual sub-spaces.

(2024) Virtual constructs (NASEM, 2024): NASEM’s recent definition of a DT focuses on a “set of virtual information constructs that mimic the structure, context, and behaviour of a natural, engineered, or social system (or system-of-systems). It is dynamically updated with data from its physical twin, has predictive capability, and informs decisions that realize value. The bidirectional interaction between the virtual and the physical is central to the digital twin.”

(2021) Royal Society (Royal Society, 2020): A DT is a virtual representation of a physical asset that can be used to understand, predict, and optimise performance. Simulations can be run before manufacture, or during use, with the possibility of (near) real-time data feedback. Feedback enables a control loop, with the possibility to adjust the real-world set-up based on insights from simulation. In turn, simulation capabilities become a high priority, throughout the system’s life cycle. DT solutions are proposed for industries such as agriculture, highlighting a growing interest in DTs that involve both natural and engineered systems. It is worth noting that DT applications in agriculture remain relatively unexplored (Jans-Singh et al., Reference Jans-Singh, Leeming, Choudhary and Girolami2020; Pylianidis et al., Reference Pylianidis, Osinga and Athanasiadis2021).

(2021) Towards Planetary Scales (Nativi et al., Reference Nativi, Mazzetti and Craglia2021): Emerging from the fields of Meteorology and Geoscience (Bauer et al., Reference Bauer, Dueben, Hoefler, Quintino, Schulthess and Wedi2021a,Reference Bauer, Stevens and Hazelegerb) the term Earth DTs (or DTs of the Earth) has been proposed within space agencies, climate research, oceanography, and others (Ye et al., Reference Ye, Du, Han, Newman, Retchless, Zou, Ham and Cai2023). Generally speaking, these are large representations of an Earth-scale process or phenomenon. Such large-scale DTs are obtained by merging continually monitored data from a variety of sources—for example, remote sensing, in-situ measurements, and synthetic data streams. Researchers tend to focus on data assimilation and ensemble forecasting techniques, which are prevalent in the associated fields. Continual updates in Earth DTs should help to reflect potentially dramatic changes associated with the highly nonlinear systems, which approximate natural phenomena. At the planetary scale, many representations are dominated by (rather than augmented with) naturally occurring systems.

Our main take is on the increase in scale of DT representations. This progression is also recognised in the Alan Turing Institute’s white paper (Bennett et al., Reference Bennett, Birkin, Ding, Duncan and Engin2023), comparing a range of relatively simple representations (a single structure or asset) to complex interacting systems (transportation networks). Another fundamental concept is simulations via the virtual representation, allowing us to probe the design space (risk-free) for improved decision-making and insights: the ultimate purpose of a DT.

Both process and data-centric procedures must be utilised to enhance interpretation (Chantry et al., Reference Chantry, Christensen, Dueben and Palmer2021). Since the DT themselves reflect systems and processes, we can only progress with deeper process understanding and the associated interactions. This is core to science in general. Intuitively, process models will be central when scaling DTs, as they inform the framework that connects DT components, which must also evolve alongside models and data. Following Blair (Reference Blair2021), we consider process modelling and data modelling as two points on a spectrum. An underlying consideration of DT design is to identify a sweet spot, whereby process and data understanding work together.

2. Three pillars

We present an alternative view of DT concepts based on three keystone features: Emulation, Evolution, and Aggregation (see Figure 1). Whilst seemingly abstract, these pillars aim to provide a general methodology to underpin the notion of a hierarchical aggregation of knowledge, an artefact emerging throughout the literature.

Figure 1. An alternative view of digital twin concepts based on three pillars: allowing extensions through aggregation.

3. A hierarchy of scales

DTs started with modest ambition—in effect, a product of engineering asset management. The vision has evolved, and as we expand from smart homes to smart cities, the knowledge encapsulated within DTs increases. We now focus on the sensing, monitoring, and control of entire engineered ecosystems. We explore the consequences of this progression in the design space, extending the ideas of hierarchy, aggregation, and federation from Section 4. First, we summarise increased expectations as a hierarchy of scales (1–4):

• 1. System-specific: Typical procedures and tools for the monitoring and simulation of a single system or asset, relating to conventional asset management, prognostics, and control (Bruynseels et al., Reference Bruynseels, Santoni and van den Hoven2018; Barricelli et al., Reference Barricelli, Casiraghi and Fogli2019; Furini et al., Reference Furini, Gaggi, Mirri, Montangero, Pelle, Poggi and Prandi2022). The boundaries of the virtual representation (or the limits of the DT) are intuitive, with clear approximations.

• 2. Local collection: Aligned with the ideas from the Internet of Things, including connected, distributed representations on a smaller scale. Examples include industrial applications (Kappel et al., Reference Kappel, Brecher, Brockmann and Koren2022), urban farming in the United Kingdom and the United States (Jans-Singh et al., Reference Jans-Singh, Leeming, Choudhary and Girolami2020), and land-based aquaculture (Lima et al., Reference Lima, Royer, Bolzonella and Pastres2022). The boundary of the representation becomes less clear—for example, in agriculture, the soil depth and surrounding environment become increasingly complex.

• 3. Ecosystems: Typical applications include urban analytics, smart infrastructure (Oughton and Russell, Reference Oughton and Russell2020), population-based structural health monitoring (Worden et al., Reference Worden, Bull, Gardner, Gosliga, Rogers, Cross, Papatheou, Lin and Dervilis2020), and Industry 5.0 applications (Braque et al., Reference Braque, De Nul and Petridis2021). More recently, large-scale applications in health, medicine, and education are emerging (Björnsson et al., Reference Björnsson, Borrebaeck and Elander2020; Furini et al., Reference Furini, Gaggi, Mirri, Montangero, Pelle, Poggi and Prandi2022). While these aggregates are typically constituted of engineered systems, with defined boundaries, the limit on the extent of the DT representation is harder to define.

• 4. Planetary: Distributed and parallel DTs combined to accomplish scaled-up tasks of complex interconnected systems (sensing for weather, pollution, large-scale energy modelling, etc.). Highly complex boundaries and interactions are associated with the aggregate.

Unsurprisingly, it becomes difficult to define large-scale DTs in plain practical terms. The concepts of aggregation (Grieves and Vickers, Reference Grieves, Vickers, Kahlen, Flumerfelt and Alves2017) are steadily increased from the combination of virtual subsystems (for a single artefact) to many nested groups of interacting models. Various issues are related to combining and coupling DTs: emergent properties, impedance matching, and uncertainty proportions. The required systems thinking will have both cultural and theoretical issues lying in wait, regarding modelling and data management. All of the above highlight a (pre-established) necessity to formalise message passing and knowledge transfer between DT components. In other words, a framework for the combination of models, data, and outputs from many related DTs.

3.2. An example of increasing scale: wind energy

Figure 2 presents one natural example of DT aggregation, progressing through the hierarchies of scale with lighter shades of purple.

Figure 2. Scaling digital twins by aggregation: wind energy systems.

• 1. System-specific: A single turbine—even at this scale, the DT is an aggregate of data + models, for example, corresponding to different components and their interactions (the nacelle, blades, and foundations).

• 2. Local collection: The wind farm of constituent turbines. The boundary becomes more complex: where is the limit of representation and how do we capture the interactions between other turbines, the grid, and the environment (on or off-shore)?

• 3. Ecosystem: Alternative power sources are now included, alongside human agents. Interoperability of DT information becomes key, especially between organisations. The boundary extends to both natural (human as an agent) and engineered system behaviour.

• 4. Planetary: Repeated extensions (into the background). The DT combines several whole energy systems; this could be international. Note that the elements of the aggregate can be shared between parts (e.g., the weather models).

4. The design space: models + data

Traditionally, machine learning researchers have adopted a model-centric mindset; their focus is on model architectures, tuning hyperparameters, or tailored loss functions to improve prediction. An alternative mindset is emerging, which places the data themselves at the centre of the philosophy—Data-centric ML (or data-centric AI; Strickland, Reference Strickland2022). We see this approach as harmonious with conventional statistical, scientific, and engineering processes. Given our process understanding, encoded in causal or scientific models (domain expertise), how do we collect data to test our hypotheses? Experimental design and data curation become a key focus at least as important as model design. We have already discussed how process understanding is critical for DTs at scale, and a data-centric mindset conveniently integrates with this philosophy to curate datasets that further our understanding of the properties and behaviours of the systems of interest.

A simple consequence of this view is that we consider the design space as an extension of the model space, as shown in Figure 3. We encompass data and deployment, as well as models/algorithms, to constitute the evolving virtual representation. Process understanding must inform the design of each aspect.

Figure 3. Data-centric machine learning: folding models, data, and process understanding into the design space.

To demonstrate, we discuss these extensions around a few important dimensions.

Modular data collection: Modular representations will combine many partitioned systems—their data are likely to be siloed and/or incomplete; however, the whole is something besides its parts (Metaphysics, Aristotle, 1908, translated by W.D. Ross). Here, a data-centric mindset has great potential, where data slicing includes cleaning, wrangling, and other data transformations in view of the collected representation rather than individuals (Data-centric AI: https://spectrum.ieee.org/andrew-ng-data-centric-ai). Aside from the push for big-data and universal models, we believe that most successful DTs will focus on curated, high-quality data and a combination of relatively specialised models to provide interpretable representations.

Encoding prior knowledge: The formulation of models must be carefully considered. This allows us to encode our prior intuition within DT representations, given knowledge of the underlying processes. Typically, this knowledge is encoded in two ways: (i) through interpretable model structures and (ii) our prior belief of the underlying distribution of unknown parameters. The prior representation is critical since it exists before any data are observed: before the model has learnt from data. With reference to Pillar 1, the prior ensures the quality of the emulation, constraining and informing the mirror, so that predictions align with reality. In a survival model example, predicting a negative age for life expectancy is clearly unreasonable: the DT representation must not permit such a prediction. The emergent areas of physics-informed machine learning (Mao et al., Reference Mao, Jagtap and Karniadakis2020) will prove critical here to adjust/constrain the model space to reflect reality. Importantly, the prior is essential to inform experimental design and data collection (implicitly or directly) as it underpins a data-centric mindset and informs our curated data. Caution is required, however. Like data, if our prior knowledge is incorrect, this can be detrimental to the DT representation.

Cyber-physical components: Of course, cloud computing (private or public) presents a natural starting point for DT models. This follows the scaling of data computation/storage, alongside various new software for parallel, distributed, and federated learning. Not only does a cloud-based residency for DT models facilitate the distribution of knowledge outside traditional data silos, it supports the necessitated scale required to achieve aggregated, planetary scale DTs. The alignment of distributed sensors to cloud-aware technologies presents an exciting opportunity. Sensors and actuators have been integrated within engineering solutions for hundreds of years, with the theory of control feedback dating back to James Clerk Maxwell (Reference Clerk Maxwell1868). As such, a distributed approach (Fuller et al., Reference Fuller, Fan, Day and Barlow2020) to data transfer, storage, and processing should naturally integrate with DTs at scale. With on-device data analytics receiving notable attention (Bagchi et al., Reference Bagchi, Siddiqui, Wood and Zhang2019) and discussed in Section 5, the opportunity to tailor data collection at the edge, around the requirements of the DT, is presented.

Another interesting trend follows smaller components and the capability to run on scavenged power. These systems do not require connection to the power grid, enabling sensing at increasingly remote locations. Additionally, the drive to make bio-degradable (or at least neutral) devices for environmental sensing is emerging (Zhu et al., Reference Zhu, Wen, Zhang, Huang, Liu and Hu2023).

Systems heterogeneity: Two concepts are important—heterogeneity of the physical twin, alongside that of the virtual. In the virtual sub-spaces, structured and parameterised models are more easily categorised into groups that inform knowledge transfer and the potential for data sharing, as similar categorisation would be useful for physical systems (Gosliga et al., Reference Gosliga, Gardner, Bull, Dervilis and Worden2021). An intuitive example returns to the wind farm, with turbines of the same make and model. These systems will have interpopulation variances—local environment, manufacturing differences, and location within the farm. Another wind farm might include alternative turbine models (e.g., analogous to a population of similar species). Despite many differences between the machine groups, these two populations should still share characteristics (or parameters)—for example, the parametrisation of their power curves (Bull et al., Reference Bull, Di Francesco, Dhada, Steinert, Lindgren, Parlikad, Duncan and Girolami2023). The design of both data and models informs the suitability and manner of DT aggregation.

Openness of the system at the edges: We can attempt to measure all aspects of a railway system, but the entire ecosystem is far too large and complex. We must draw boundaries to constrain both measurement and modelling, rather than modelling alone. For example, if a DT estimates the response of a bridge from train loading, how do we model the varying mass of commuter trains? Do we average, or model passenger dynamics at rush hour? We see that boundary definitions become increasingly difficult as DTs include naturally occurring systems, that is, human behaviour or weather/ocean models (UN DITTO programme on DTs of the Ocean https://ditto-oceandecade.org/). Another consideration of boundaries is purely virtual, concerning legacy software; for example, this could be resolved with wrappers for archived code in evolutionary architectures.

Levels of autonomy: One aim of DTs is to reduce the reliance (and overspending) on human interactions. Progress will require investigations of control and decision-making informed by data (Langtry et al., Reference Langtry, Zhang, Ward, Makasis, Kreitmair, Conti, Di Francesco and Choudhary2023), likely towards more AI-based control, such as reinforcement learning (Cronrath et al., Reference Cronrath, Aderiani and Lennartson2019). By guiding interactions and data collection with the virtual representation, we are once again (quite naturally) within a data-centric ML mindset: whereby models, actions, and data collection evolve alongside their physical counterparts. Autonomy implies decreasing levels of human input; however, the rate at which human interactions are replaced will depend on the type of action. For example, the formulation of domain expertise will be very difficult to automate (and illogical) in most settings.

An extensive system design challenge exists for a scalable DT architecture. From a purely technical standpoint, we propose that systems researchers should consider a clean-slate approach to scale distributed processing. That said, considerable investment in the current systems could make any new architecture financially prohibitive. Discussions would be similar to those surrounding new Internet architectures (Feamster et al., Reference Feamster, Rexford and Zegura2014).

5. Beyond data privacy

Privacy must go beyond the data themselves to encompass the evolving aspect of aggregate models. While the data typically come from critical infrastructure or commercial services, they should not be the only concern regarding privacy. As DTs develop, we require a new type of governance to ensure secure knowledge exchange from sensing to deployment. Most obviously, this includes both data and modelling components of continually evolving representations. During DT design, it will be important to incorporate privacy-enhancing technologies, both hardware and software, to protect the models and data within the core system. The approach should minimize the possession of personal data (Bruynseels et al., Reference Bruynseels, Santoni and van den Hoven2018) at any of the four stages of the hierarchy of scales. Whilst the sophistication of knowledge increases at each scale—through the aggregation of previous levels of knowledge and data—so does the repercussion of any privacy implications due to residual personal data. We discuss privacy considerations while progressing through the hierarchy, illustrated in Figure 4.

Figure 4. Privacy concerns when scaling digital twins—smart energy meters.

6. Ethical considerations

The proposed three pillars of DT representation rely on data pathways to update the mirror (emulator) evolving through time. We can then virtually trial what-if scenarios (Schluse et al., Reference Schluse, Priggemeyer, Atorf and Rossmann2018), investigating how alternative interactions might influence real-world processes and avoiding costly in-situ experiments. While the experiments are virtual, the results inform real-world decisions. Helbing and Argota Sanchez-Vaquerizo (Reference Helbing and Argota Sanchez-Vaquerizo2022) have recently discussed the ethical implications of viewing DTs in this way as duplicates (or imitations) of reality that interfere with our individual thoughts, decisions, and behaviour. Consider a refocus towards humans, rather than automation, in Industry 5.0 (Commission et al., Reference Breque, De Nul and Petridis2021). As we bring humans back-in-the-loop critical concerns relating to human rights and dignity are raised. That is, we expect to interact within increasingly virtual spaces (Commission et al., Reference Breque, De Nul and Petridis2021). Importantly, since the outputs of any virtual representation can have a very real effect on human experience, they must be subject to ethical scrutiny. As we aggregate DTs, the concerns of misuse progress from individual (or corporate) to societal. To aid discussion, we adopt the considerations of Helbing and Argota Sanchez-Vaquerizo (Reference Helbing and Argota Sanchez-Vaquerizo2022) and Caldarelli et al. (Reference Caldarelli, Arcaute, Barthelemy and Helbing2023) to reflect on some ethical design issues (EDIs) when scaling DTs.

7. Concluding remarks

Rather than redefine, we have gathered emerging concepts in the DT literature—highlighting requirements and areas that should prove critical when moving forward. In doing so, we have identified several important themes: model/data aggregation, design space boundaries, privacy, and ethical concerns. While most researchers recognise these elements as crucial, they are usually discussed separately, across different communities, and rarely considered holistically. The backbone of our discussion follows a hierarchy of scales that correspond to increasingly aggregated units. Following this modular approach, the focus when collecting DTs is on the framework of connections, rather than the models themselves. As such, local expert knowledge (and ownership) of the real-world system can stay connected to local DT units, while the aggregate is naturally decentralised, with shared control. Despite a modular structure, the whole can be iteratively refined through the pillars of emulation and evolution, as discussed. This mindset of DT representation naturally addresses concerns throughout many themes from practical implementation and sensing to ethics and privacy. By embracing these ideas, the vision of an interconnected DT ecosystem becomes more attainable.