Digital twin maturity levels: a theoretical framework for defining capabilities and goals in the life and environmental sciences

Digital revolution

With the advent of concepts such as the Internet of Things (IoT) computers can now more readily ‘sense’ their environment in (near) real-time using a variety of embedded sensors all networked together.^1,2 Many of these IoT devices are everyday items,¹ others are small increasing their portability and ensuring that where there is the possibility of a connection to the internet, or another network, observations can be made.² One application of such interconnected sensors is the acquisition of data from an asset that maybe too big (e.g., a city), remote (e.g., streetlights), or inaccessible (e.g., inside an air conditioning system) to study in detail with traditional data collection methods. Significantly, applying these interconnected sensors to a problem allows for the collection or generation of a continuous stream of data independent of analysts allowing for higher resolution, continuous datasets to be acquired. These developments are considered to have significantly furthered our progress to the next, fourth, iteration in our industrial development.^3,4

The fourth industrial revolution (Industry 4.0 or 4IR) should see the transition from the digital revolution⁴ - represented by the switch from mechanical and analogue systems to automated digital electronics - to industrial processes dominated by cyber-physical systems, smart sensors, and cognitive computing. Conceptually the fourth industrial revolution is built upon expanding the digital technology revolution^5,6 that occurred during the third industrial revolution through increased interaction and integration^7,8 between the virtual and physical (i.e., “real world”) environments. Such interaction and integration can take many forms (Figure 1), such as (digital) models that simulate processes that occur in the physical world; digital shadows or digital threads that collect, store, and provide information about a physical object; or digital generators, devices that construct physical objects (e.g., 3D printers).

Figure 1. The current ’Digital’ landscape.

Variants of digital technologies characterized by how information flows between the real (blue) and virtual (green) worlds. Examples of these varieties include 3D printers and barcodes here referred to as a digital generator and digital shadow respectively. Solid black lines denote an automated flow of information, whereas grey lines denote a ‘manual’ flow of information. Based upon Refs. 7, 8.

Digital twins

Associated with the coupling of the virtual and physical into cyber-physical systems is the potential formation of digital twins^9–11 with commercial^12–15 and academic^16,17 interest piqued regarding the possible application of digital twins to various complex processes and systems (e.g., Ref. 18). Digital twins can be differentiated from other ‘digital’ variants (Figure 1) by the connection between the virtual and physical realms having a continuous automated two-way flow of information¹⁶ (Figure 2). Through this communication, a virtual representation simulates how a process or physical asset operates throughout its lifecycle^19–21 from which decisions can be made that control the real-world process. Such simulations can be simple, for example keeping up to date readings of sensors that register the status of the physical asset. Or the virtual simulations can be anticipatory, using prior knowledge, time-series, and/or physics-based models to determine the likely future outcome based upon the current status of the physical asset. The feasibility of creating digital twins has existed for a time in the manufacturing and engineering sectors, where the intricacies of components and their complex processes are known in detail or can be intuited from blueprints.

Figure 2. Concept and components of a digital twin.

(A) Conceptual components of a digital twin, having a (1) physical asset; (2) virtual representation; (3) two-way communication; and (4) feedback or actionable knowledge. (B) In reality the various components maybe more complex. The terms used in (A) to describe the various stages, i.e., ’Collects’, ’Processes’, ’Discovers’, and ’Manages’, are from Ref. 12.

From such information it is possible to place sensors that can generate data of sufficient quality and resolution to allow for the development of Digital twins that can monitor, control, and optimize processes. Or remove geographic and temporal constraints, so that operators can be trained remotely or decision maker can view historic, current and future states of the physical asset whilst not needing to be ‘on location’ or proximal to it.^22,23

Digital twins in use

Beyond the factory floor the digital revolution is taking place in industries related to the Life and Environmental Sciences.¹⁷ Such as Utilities,¹⁸ the Agricultural sector,^24–28 Horticulture,²⁹ Apiculture,³⁰ Smart farming,²⁴ Animal farming,³¹ as well as the Healthcare sector,^32,33 and the Pharmaceutical Industries³⁴ amongst others. Other twins aim to provide information on the built-environment (e.g., planning, environmental impact, air quality, etc.), advice on sustainable water management [1], or to help protect and restore (European) biodiversity [2]. More ambitious digital twins aim to provide a digital twin of the Oceans [3] and our Planet [4] to assist in the green transition.³⁵

Such an interest in the possible application of Digital Twins to the Life and Environmental sciences has been ignited by the emergence of new techniques and new technology that can circumvent the complexity of real-world phenomenon. New techniques include those that can derive patterns and connections without the explicit requirement to know the process behind such patterns (e.g., machine learning) or those that enable a model to be kept synchronised with observations (e.g., data assimilation). Alongside the development³⁶ of smaller more portable interconnected technology and devices,³⁷ loaded with sensors^38–40 that can acquire sufficient data to feed into the virtual half of the twin.

Yet given their integration into various aspects of our daily lives’ digital twins for the life and environmental sciences may have the potential to cause considerable socioeconomic upheaval. Such upheaval may occur through changing our interaction with ourselves with the development of personalised healthcare^38,40–42 and (personal) environmental monitoring. Or via adapting environments (e.g., greenhouses) and organisms to meet our needs e.g., optimisation of agricultural practices. However, for all the possible benefits and opportunities of utilising digital twins for life science applications there are many barriers to success such as the ‘blueprints’ of a natural process or phenomenon being only partially known, too complex, technically unfeasible or unethical to be obtained. There is also the potential for such socioeconomic upheaval to exacerbate rather than curtail existing problems, such as those relating to social injustices or healthcare^43–46 or ensuring continuation of poor practices (e.g., in agriculture⁴⁷).

Digital twin definition

There is no explicit formal definition of a digital twin leading to some divergence between various (scientific) disciplines (e.g., Ref. 50) about what a digital twin actually is. To further exacerbate the problem studies have shown that in some papers what is referred to as a digital twin has not always been clarified⁵¹ leading others to attempt to catalogue the variation in descriptions^{3,7,16,24,52–54} including identifying what can or cannot be considered a digital twin.^7,8

From a practical perspective here, it is considered that a definition of a digital twin is valid if it meets the following criteria of being a virtual representation, replica, or simulation of an actual, or possible, physical asset in (near) continuous two-way communication with its counterpart to play an active role in planning, decision, and control processes through providing feedback. The digital twin will be primarily data-driven and capable of, or have the capacity for, predicting future trends or developments over a period of time. Here feedback need not be an automated process but simply incorporation of the results of the twin in a decision-making process. There are four commonalities between the various types of digital twins: (1) a physical asset; (2) a virtual representation; (3) two-way communication; and (4) some form of feedback (Figure 2). In the following, let us expand upon these various components.

Physical asset

A digital twin must be comprised of a physical asset in physical space and a virtual replica, copy, analogy, or representation in virtual space. There is no exclusivity with respect to twinning, so the same physical asset can be twinned by multiple digital twins.⁵⁵ Likewise, the same sensors could be used for multiple twins. In some descriptions of digital twins whether they are a part or the whole asset has been differentiated by adding the following terms ‘component’, ‘asset’, ‘system or unit’, or ‘process’ to the term digital twins.^15,56 Where a ‘component’ is part of an asset and a ‘system’, ‘unit’, and ‘process’ refer to a greater and greater collection of assets. Therefore, in digital twinning the physical asset need not be the ‘whole’ instead it can be a singular component of a larger system (e.g., a motor in a factory; the digestive tract of an animal; etc) or alternatively a group of individuals (e.g., a fleet of delivery vehicles; a herd of cows; a field; etc.) as long as it can conceivably or does physically exist. Similarly, digital twins that describe a prototype of a physical asset, explicitly one that can exist, have been referred to as a Digital Twin Prototype to differentiate it from those with a real-world physical asset. In the life sciences the digital twin prototype can possibly occur in the biotech industry representing organisms in a natural state or as genetically modified organisms, where the cells may be modelled first prior to their creation.

Virtual representation

In the virtual environment, the physical asset is twinned by a virtual representation that should in theory provide the user with relevant information they could obtain were they to go out and physically inspect the asset in the real-world. Assuming that measurements, time, money, and distance are not limiting factors. Of course, in reality the twin utilises a subset of measurement components that either directly, or indirectly (e.g., proxies) capture various states of the measured system. The virtual representation can be broadly split into a visualisation component (what the user sees) and a processing component (how the data is transformed, used, or incorporated into models). Some definitions of digital twins implicitly state that in order to have maximum fidelity with the real-world asset by necessity the visualisation component of virtual representation must be a 3D model of the asset. Here, however, the requirement of the digital twin to have a 3D visualisation is omitted as ultimately the visualisation should depend upon the (i) type of data; (ii) the goal of the visualisation; and (iii) the needs, wants, and requirements of the user. Hence, the visualisation components can range from simple iconography, a dashboard, graphs and graphics, a 3D model, augmented reality or virtual reality. Whereas the processing components can range from just the data, a data-model, physics-based model, artificial intelligence, or some combination of all of these. Ultimately the form the virtual representation takes will be dependent upon the knowledge of the targeted user group and the purpose of the digital twin.

Two-way communication

Between the virtual-physical and physical-virtual spaces two-way communication should occur (Figure 1). Of course, the nature and method of this two-way communication will vary. Communication between the physical and virtual worlds may be via networked sensors, data upload, or manual data-entry (e.g., patient results). Communication between the virtual and physical worlds could be in the form a simple visualisation (e.g., icon forms or composite interactive dashboard) or more complex feedbacks and actions (e.g., automation).

Likewise, the frequency of communication will also vary being dependent upon the use case. For instance, high frequency communication may be required where rapid changes occur but just as equally when trying to pinpoint subtle variations. It may be prudent to first identify the physical assets’ own unique trends and rhythms prior to defining the communication interval to ensure the right frequency of communication occurs. For example, if one was aiming to twin a plant then consideration must be made regarding the frequency of communication so as to sufficiently capture the various internal and external natural trends and rhythms (e.g., diurnal or seasonal cycles). In other instances, there may be variables or components of the Digital Twin that are updated only infrequently. For example, when representing a complete city, the volume of traffic might be updated continuously whereas road layout itself is semi-rigid - being fixed but requiring updates at a vastly different time scale and often performed manually by cartographers.

Actionable knowledge

Finally, there must be an initiation of a decision-making process and where required feedback; with the Digital Twin providing the source of actionable knowledge to aid in the decision-making process. Actionable knowledge refers to the coupling of Science and Management.⁵⁷ That is using directed action to bring about intended consequences (i.e., management) whose actions and consequences are defined by the generalisations provided by empirical research to describe or explain a particular phenomenon (i.e., science). Actionable knowledge will be knowledge that can provide either the specific conditions or actions necessary to ensure or lead to a particular set of intended consequences beyond the settings in which it was first created, and which can be applied to both individuals as well as to groups.

Digital twin

In summary, a digital twin is composed of a virtual representation or replica of an actual, or possible, physical asset with these two counterparts in continuous communication so as to play an active role in planning, decision, and control processes. Following coupling between the physical and virtual twins both twins are collectively referred to as a Digital Twin Instance.^11,16

Going beyond a definition

As can be seen from the proceeding section a definition does not inform the reader of what a digital twin will do for a given problem.⁵⁸ A concept, or definition, of Digital Twins will only provide a ‘foundation level’ knowledge, i.e., knowing the definition of a digital twin serves to differentiate it from other types of digital technology (Figure 1) and what components are key (Figure 2A) but not much else. This is because for each criterion (i.e., virtual asset; two-way communication; feedback; and a decision-making process) there can be a range of potential outcomes (Figure 2B) giving an equally wide variety of possible digital twin end products, or more aptly ‘phenotypes’ (i.e., observable characteristics or traits). How the various components of a Digital Twin are expressed and formulated must be discussed during the development phase. The resultant Digital Twin will depend upon the end goal of the digital twin, as well as any legal and technological practicalities, and the various decisions made in the design phase.

Solving the question of what a particularly digital twin will do and its capabilities necessitates communication between developers, stakeholders, and users. Yet, communication between stakeholders requires a common frame of reference. Hence, in order to solve the issue of what a digital twin will do, several initiatives have placed digital twins onto a hierarchical framework (Figure 3), describing various types, capabilities, and the evolution^59–63 or developmental stage²⁶ of digital twins.⁶⁴ These have been referred to as digital twin maturity levels,^59,65,66 a series of digital twin integration levels,⁷ a continuum,⁶⁰ a spectrum,^65,67 archetypes,⁶⁸ and/or a classification⁶² or taxonomy of digital twins,⁶⁹ all of which seek to define the functionality of the various digital twin levels (Figure 3).

Figure 3. Synthesis of digital twin maturity levels.

Comparison of the various levels of maturity^{59–63,65,70,71} placed upon a relative scale where the sizes of the boxes reflects the coarseness of the granularity. Blue boxes represent levels that do not represent a Digital Twin. Whereas, dashed unfilled boxes represent levels not specifically defined by the schema. References with a * are from online (business) blogs, ** from technical reports, and others from academic journals.

Digital twin maturity

Previous work

As such investigating previous work is a useful starting point for defining a digital twin maturity model that is applicable to Life Science problems. However, several previous studies have expanded their models to incorporate other digital instances (Figure 1). For example, Kritzinger et al.’s⁷ digital twin integration levels put digital twins on a continuum involving digital models and digital shadows (Figure 1). For brevity only previous work differentiating between different types of digital twins will be discussed here. Where the different types of digital twin reflect differences in functionality of the digital twin. As opposed to the type of asset being twinned (i.e., Refs.15, 56).

A quick review of this subset of the literature (Figure 3) suggests that the types of digital twin are generally differentiated by the level of insight that the twin provides about a problem or toward a goal.⁷¹ A digital twin having better insight can be seen as having increased capabilities alongside an increased capacity to perform the actions itself. It is our opinion that Digital twins can be broadly categorised into those that provide ‘Data about’, ‘Knowledge of’, ‘Suggestions for’ and ‘Control of’ a physical component, asset, process, or phenomenon. These can be split into data collection, analysis, prediction, optimisation, and autonomy (Figure 3). Differences between previous studies largely reflects the ‘coarseness’ of the various digital twin sub-divisions. For instance, ‘twins with business models’ reflecting a stage where the digital twin focuses upon optimising a process⁵⁹ can itself be sub-divided into various substages: an initial substage coupling the twin to business operations, a subsequent substage recommending various decisions,⁶⁰ and finally making those decisions autonomously.⁶³ Yet despite differences between the assorted maturity models there is a general consensus that to reach the foundational level of a digital twin requires an availability of data⁷¹ and/or a connection between the physical and virtual worlds.^65,70 The lowest level of a digital twin is generally reserved for a twin that simply relays the collected data, or real-time status, and the highest level when the twin has gained autonomy to control.⁷² With many prior maturity models referencing automobile automation or the SAE’s levels of driving automation for self-driving cars⁷³ as a comparable frame of reference (e.g., Ref. 59). Between these two end members (data collection and automation) there are levels devoted to analysis (including modelling, simulating), prediction, and optimising (Figure 3).

Conceptual framework

For the framework outlined here we continue the use of the term maturity^59,65,66 given how it denotes both a natural, and sequential, progression toward more complex states (e.g., see the data-driven manufacturing maturity model⁷⁴). Although it is entirely feasible to also refer to it in terms of functionality or capability. Likewise, our maturity levels are also based on the presumption that the ultimate goal of digital twins is to streamline decision making processes through the automation of sequential tasks. Whether or not the ultimate goal of an autonomous twin is obtainable is another matter (see, Discussion) and there is some debate regarding digital twins and cyber-physical systems (e.g., Ref. 53). In our framework the various maturity levels can be conceptualised as a series of progressive stages⁷⁴ that add functionality towards the goal of automation.⁷³ At low digital twin maturity levels, the user or operator of the digital twin carries out more of the data capture, decision making, process and control operations. As a digital twin matures processes once controlled by human operators become subsumed by the virtual twin, until the process is fully automated. These levels focus upon providing insight upon a particular aspect of the object under study, with aspects that address the what’s, the how’s, and the why’s. What questions relate to conveying the basic research elements, (raw) data and information (‘what is the time’, ‘what value does X have’). How questions relate to understanding the relations of these basic research elements, e.g., through models, from which predictions can be inferred. Why questions relate to identifying cause and effect as well as whether something can or should be done. Why can be sub-divided into ‘appreciation of why’ and ‘why’. This sub-division is based upon the idea that there are instances where the reasoning or knowledge behind why something occurs may not be known (’why’) but an answer or explanation can still be formulated (‘appreciation of why’).

To explain the conceptualisation let us consider a simple example, that of traffic in a city. If we want to know the amount of traffic on the road, we can begin by asking a simple question, “What is the number of vehicles on the road?”. Such a question relates solely to the measurement of vehicles on the road (i.e., “6 vehicles” or “5 cars, 1 bus, $\dots$”) and it is up to the receiver of that data to provide the context. By changing our question to “What is the current volume of traffic?” context and meaning are added to the answer (i.e., “low”, “high”). But it might be prudent to know “How will traffic change over the course of a day, week, or month?” and therefore, understanding the variables necessary for prediction and modelling becomes necessary. Knowing the amount of traffic can be used to provide insight for a practical use, for example selecting the optimum route: “What routes are there from location to A to destination B?”. Finally, such knowledge could be used to address “What is the best route and best time to leave from A to get to B?”

The basis of this conceptualisation is the knowledge pyramid (or knowledge hierarchy; information hierarchy; knowledge management pyramid),^75–77 a common feature in information and knowledge management.⁷⁷ The knowledge pyramid is represented by a hierarchy commonly presented as a pyramid proceeding from the base of data to information, knowledge, understanding, and ultimately the pinnacle representing wisdom. To transition from data to information requires understanding relations; information to knowledge requires understanding patterns; and knowledge to wisdom requires understanding principles. The stages of the pyramid hierarchy can be mapped to the conceptualisation’s different questions, the data and information relating to ‘knowing what’, knowledge to ‘knowing how’, and wisdom to ‘knowing why’. Understanding is linked to ‘appreciating why’, the why might not be fully understood but cause and effect might be. The mapping between the hierarchy and these concepts has been argued by Zeleny⁷⁶ who stated that to “manage wisely implies knowing why to do something; to manage effectively implies knowing what to do; to manage efficiently implies knowing how to do it (and to ‘muddle through’ implies nothing and having ‘lots of data’ around)” [5].

Digital twin development

Digital twins are being developed in a variety of disciplines both commercially and academically.^{17,18,24–34} Therefore, these projects may encompass a wide range of use cases and stakeholders who may not even have a cursory knowledge of what a digital twin is, let alone an understanding of what it will do. Likewise, there won’t be a ‘one-size-fits-all’ approach to developing digital twins (as outlined in the ‘Digital twin definition’ and ‘Going beyond a definition’ sections above). Therefore, as previously noted, development of a digital twin boils down to two fundamental methodological questions. The first question relates to the full range of the digital twins features and possible actions, i.e., ‘What will the digital twins capabilities be?’. The second aims to address which of the capabilities identified in the initial question the digital twin will it utilise, i.e. ‘What will the digital twin do?’. The answer to these questions can be considerably broad, especially depending upon the intricacies and peculiarities of the use case (e.g. note the range of expressions in Ref. 17). Critically it is important that in the process of describing what a digital twin will do that we end up with less of a wish list and more of a list of requirements for development. The levels of digital twin maturity proposed here aims to serve as a means for effective communication between project leaders, project owners, project participants, developers, and stakeholders about the intention of the project. Allowing developers to outline to interested parties what the goal and purpose of the digital twin will be. The maturity model should also act as a springboard in defining what is feasible with current resources; and, to know where a digital twin is in its development phase.

Example: Twinning a house plant

As a simple example, let us consider how the maturity levels could be applied to twinning something simple such as a house plant (Figure 2). The first step in digital twin development will be to perform an initial scoping exercise that will preliminary determine what type of digital twin (Figure 4) suits the intended objective.⁵⁹ This exercise should identify the asset under study; an intended goal; prior knowledge; measurable and required variables; define processes; determine the expected values of key variables; define the intended users and stakeholders; and consider if and what output and (anticipated) feedback the digital twin will produce. For a house plant (our object) there are several parameters that can be directly affected by an individual (our stakeholder) to promote growth (our goal). These include providing sufficient light, water, and temperature, all of which can be measured to a greater or less degree of accuracy by commercially available sensors. Once the initial scoping exercise has been performed, a digital twin developer should use the answers obtained to sketch out the rough divisions between the various maturity levels. Using coarser divisions such as data collection and communication (ML 1-2); analytics, models, and predictive tools (ML3); and ‘use cases’ (ML 4-5) can help make preliminary design decisions. Following which, it is prudent to next identify previous or preliminary work that already fulfils or can feed into one or more of the maturity levels. Applying the maturity model follows an iterative design methodology with each level, or iteration, being a cycle of development, testing, analysis, and reflection. The result of each learning cycle can then be used in subsequent design iterations, either another iteration at the same maturity level or progression onward.

Assuming that the process starts from the initial maturity level (ML1), design choices should address what sensors will be used and how these sensors will communicate their data to the virtual twin. For example, questions to be addressed include ‘What is the nature of the communication?’ e.g., WiFi, LoRa, (etc.), ‘How frequent will these sensors communicate?’ or ‘Will data arrive as a stream or be sent in batches?’ amongst others. This level adds low level knowledge of a process, such as insight into range and distribution of obtainable values and/or how such variables change spatially or temporally. Such knowledge may help develop understanding of a process. Returning to our example, a digital twin for a house plant at ML1 provides real-time data capture of the amount of light, the soil moisture, and the room temperature. Values can be collected by an IoT sensor and presented on a display next to the plant, or as values on a simple web server.

Having obtained and communicated data in ML1, at the next maturity level, ML2, such data will be given context. Here ‘context’ is conveying known information about the subject matter, either from prior knowledge or previous data. Defining which context will be added requires communicating with experts, users, or a combination of both to identify key parameters. For instance, for our house plant, context could be added through a knowledge rule that translates values, such as the light into ‘low’, the soil moisture into ‘dry’, or the temperature into ‘cold’. One could colour code the values on the display or add simple iconography to indicate (un)suitable conditions for the plant, allowing someone to identify problems and come up with their own actions. Such as moving the plant to a sunnier or shadier position, to water it, or to adjust the thermostat.

Progressing to the next level requires the twin to shift from solely focusing on past and present values to predicting future values. Hence, ML3, involves adding predictive capacity. For our example that could be knowing the future value of soil moisture and when it will drop below a specific level that requires watering. Generally, predictions will be made via a model, though the exact type (e.g., mechanistic, process, physics based, machine learning, etc.) will depend the nature of the digital twin (e.g., the use case; availability of data; etc.). Though developers can utilise the previous levels to identify required inputs and outputs as well as those variables that are of importance, those that vary, and which could be conceivably parameterised. For instance, for our house plant the digital twin could incorporate a mechanistic model that calculates the length of day and the angle of the Sun by giving the latitude, date, and time which can then be used to constrain the upper or lower light levels depending on weather. Whereas, predicting soil moisture and/or temperature could make use of a time series machine learning algorithm to make predictions without direct inference of the cause of the variation. However, crucially the model(s) must allow for the updating of various states. Techniques such as data assimilation can be used to combine new observations and model forecasts to update the prediction.

The final levels (ML4 and ML5) transition from relaying knowledge (ML1 to ML3) to creating and using new knowledge, with these levels being split between optimisation (ML4) and automation (ML5). How either of these levels will be ultimately implemented by a developer will be highly dependent upon the needs and requirements of users, experts, and stakeholders as well as taking into account various practicalities. The distinction between ML3 and ML4 is the addition of potential corrective actions and their outcomes. In our house plant example, optimisation could focus on resolving questions such as: ‘When is the optimal moment to water to maintain constant soil moisture levels?’ or’Is it better to water when the room is cooler?’. How those questions and their answers are presented will again depend on project specifics, for example they could be embedded in a user interface or developers could allow a user to test them out in a sandbox environment. Once it is possible to optimise a process, then a natural progression is the possible addition of some degree of automation, i.e., bypassing the decision maker to quickly and consistently provide a corrective action. Therefore, at the final maturity level, ML5, the process is autonomous. For our example, as the moisture levels drops the digital twin would react by watering the plant to sustain an optimal amount of soil moisture.

Choosing the right twin

Whilst our example discusses the full spectrum of maturity levels projects developing digital twins should be aiming for and settling upon the right maturity level for a particular use case rather than reaching for higher maturity levels. This might seem far from aspirational however, the goal should be to have the best fit of digital twin to a particular goal or problem. Nor is it the intention of the maturity levels presented here is to produce a ranking system. As it is more than likely that specific digital twin use cases will have an upper maturity level limit that reflects the legal and ethical considerations as well as technological feasibility. For instance, for a given problem the highest level of a Digital Twin is dependent upon consensus on the expected values a set of variables should have and the certainty of relevant knowledge (Figure 5 based upon concepts in Refs. 59, 78–83). In situations where there is a lack of both knowledge and certainty of the values, a Status Digital Twin could provide the necessary foundational knowledge. Where there is consensus on the values and high degree of certainty regarding our knowledge of a process, then an Autonomous Digital Twin might be more prudent.

Figure 5. Selecting the right twin.

Which twin is for you? Comparing certainty of knowledge verses consensus of values. Deciding on which digital twin is required for a given problem also necessitates a critical look upon known knowns, known unknowns and unknown unknowns. In this example the initial digital twin is dependent upon two variables, whether consensus is reached on the values a variable should have allowing for knowledge rules to be produced and the certainty of relevant knowledge. Based upon ideas and concepts relating to digital twins and uncertainty communication described elsewhere.^59,78–83

Technological readiness

Our house plant example shows that the capabilities of a digital twin will however, be constrained by the availability and feasibility of technology. Automating the watering of a house plant is far more feasible than automating the movement of a plant to track light. Therefore, in this case the more automated levels will focus upon solutions that are feasible. However, it should be noted that the maturity levels outlined here do not directly refer to the technological readiness of Digital Twins or their constituent components. Technological readiness being defined by the technical progress or maturity of a technology under evaluation, a technology is ranked along a technology readiness level (TRL) scaled from research is beginning (= 1) to full scale deployment (= 9). For digital twins the feasibility of obtaining each of the maturity levels defined here will depend on the individual project, its intended goal, and thus the technology associated with it. In some instances, it is conceivable that the technology associated with higher maturity levels (e.g., machine learning models, control systems, etc.) are more ready than the technology associated with lower maturity levels (e.g., automated data capture and ingestion). Where automated or networked sensors do not yet exist for example, lower maturity levels that rely on data collection may not be obtainable but higher maturity levels such as modelling (using other or legacy datasets) may not be so constrained. Therefore, a mapping between maturity and technological readiness remains highly dependent on the topic of the digital twin. Although, it is the authors opinion that the maturity levels outlined here and the TRL are complimentary. The TRL will reflect the state of technical innovation^7,54,84 and key enabling technologies.^85,86 By using the maturity index to first identify and communicate about the type of digital twin required for a project and its goals, developers can use the TRL at the next phase of project development to identify technological feasibility and/or bottlenecks.