Abstract
The digital revolution continually spawns new terms which become iconic phrases and clichés, often only lasting for a short period of time. In the last 10 years, terms such as cloud computing, platforms, big data, smart cities, machine learning, artificial intelligence (of the ‘weak’ variety), and so on have been coined to describe recent fashions in computation and communication driving the automation of society ever further. The latest term to be added to this arsenal of clichés is the ‘digital twin’ which has suddenly taken on a new lease of life, being coined almost 20 years ago but only finding widespread favour much more recently as digital infrastructure becomes ever more embedded in our industries, cities and communities.
In strict terms, a digital twin is a mirror image of a physical process that is articulated alongside the process in question, usually matching exactly the operation of the physical process which takes place in real time. It was first coined in the early 2000s by Michael Grieves (2014) whose expertise in product design initially rooted the concept in production engineering. Since its inception, however, the concept has broadened and loosened somewhat in that it is now being applied, or rather used, to characterize a variety of digital simulation models that run alongside real-time processes that pertain to social and economic systems as well as physical systems. In this sense, an immediate question is to define the differences between a real system and any computer model of that system, and in this context, it is worth noting the conundrums that this raises with respect to the arguments about what constitutes a digital twin with respect to a city.
In general, any system that mirrors the operation of another, different system is usually defined as a model, which in turn is an abstraction from the structure and processes that define the system to which it is being aligned or compared. Models are, by definition, simplifications of the real thing and in that sense, do not aim to replicate the original system in the same detail as that system. In short, key features of the original system are thrown away and usually the model abstracts that which the scientist considers key to the features of the real system that are under scrutiny. Basically, models do not attempt to mirror everything about a real system and it is this that makes the model intrinsically different from the original system. In fact, another difference is that the physical representation of any system and its models are usually different, although this distinction is not as clear as it might appear: in other words, there may be a succession of models of a system that are composed of the same or different media as the basic system. In this sense, one might have a digital model of a real system that may be constituted in somewhat different form but also as a digital model.
It thus appears that, in general, a computer model of a physical system can never be the basis of a digital twin for many elements of the real system are ignored in any such abstraction. However, there is no doubt that some models are closer to the real thing than others, with the whole panoply of models ranging from ‘thought experiments’ which are entirely conceptual to closely tailored digital representations that attempt to mirror as many features of the real system as possible. It is even possible to conceive of a transformation of a digital model from an entirely abstract conception to a full mirror image of the system in question. If, however, the model is a complete mirror image, which is the assumed definition of a digital twin, then it might be argued that the digital twin is no longer separate from the system, but in fact is the system itself. In this sense, all physical systems can have a digital equivalent which converges and merges with the system in question. In this sense, a true digital twin running in real time is no different from the system itself and this poses the question as to how the digital twin can be used to learn about the system and used to explore, simulate and test new designs if it is the system itself. For the digital twin to be used in this way, then presumably it has to be disconnected from the real system. Then, there is a logical difficulty in doing this in that both systems will run alongside one another and if they are mirror images of one another, the question becomes ‘how can the digital twin be used to explore and inform the original twin’.
Let us make this problem a little clearer with respect to cities. The idea of the digital twin in this context has emerged from the representation of the city in terms of its physical assets. Geographic information systems, their scaling down to the level of buildings and their extension to deal with the operation of buildings in terms of energy, materials use, and maintenance using building information models software, are providing the context for extensive digital representations that scale to the level of all the physical assets in the city. Quite clearly, such systems are models of one kind in that they represent the city in digital rather than material form and may be very close to the basic physical equivalents that make up the city. But they rarely include any of the processes that determine how the city works in terms of its social and economic functions. 3D virtual models – even if they have embedded within them real-time processes, such as traffic and energy flow – are only representations that function over short periods of time and are often simply representations of the city at a cross-section in time. In this sense, a digital twin is much more like a conventional computer model in that it abstracts only a limited set of variables and processes.
Wildfire (2018) makes the very useful distinction between models (or digital twins) that pertain to what we might call the high-frequency city in contrast to the low frequency. High-frequency cities operate in real time at the level of our own personal time frames, second by second, minute by minute up to cycles of days and months, while low-frequency cities operate over years, decades, centuries, eons even. In this sense, we build different models to explore very short time horizons – what Wildfire (2018) calls ‘reactive’ models where ‘feedback and visualisations enhance real-time or near real-time interventions and improve the smooth day-to-day running of the city or asset’ and ‘predictive’ models where ‘accurate input data is used to improve longer term scenario planning to steer appropriate (and equitable) investment decisions’. In fact, I have used the term model rather than digital twin because in both contexts, the model of a digital twin needs to be decoupled from the original system if we are to use the model to inform our maintenance and/or design for the future operation of the real system. In short, we need to run the digital twin offline in some way so that we can use it to explore how to improve the real system. It is most unlikely that any of these kinds of models can be run in real time, thus matching exactly the processes operating the real system. The digital twin must always receive input from the real system if only to provide some sort of diagnosis of faults in the original system, and in this sense, there is latency involved. In the case of the high and low frequency time horizons noted here, the digital twin is likely to be nowhere near the time scale of the original system, and for it to be useful in design and planning, it needs to run in an entirely different context from the real system.
The question to be asked of course is whether or not there is any digital twin that can be run in real time alongside the basic real-time system, and if so whether or not information from the digital twin can be used in some sense to steer the basic system. In our group, we have data which is generated in real time from various railway operations and this is generated in such a way that the latency between when this is received and when the users closest to the system can use it is very small. In fact, operators looking at the trains are able to correct manually the operations of the trains, but there is still a noticeable lag between any problem generated on the basic system and any measure of response to resolve the problem. This lag is often short enough to enable good corrective action, and digital twins of various sorts are possible to enable this. It is even conceivable that the demand for trains and their supply could be coordinated digitally from linking automated payment systems which track travellers and link them to trains. But in such a case, the real system which we have been treating so far as analogue is in fact becoming a mixture of digital and analogue, the digital twin being thus woven into the analogue. The closer one gets to the real system with a digital representation, the more the twins merge to become one. Thus, the idea of a digital twin as a ‘a digital replica of physical assets (physical twin), processes and systems that can be used for various purposes’ (Wikipedia, 2018) must always be qualified: it is more likely that digital twins are not identical twins and the notion of an exact mirror is an idealization that will never be achieved.
This does not mean that the idea of a digital twin is irrelevant: far from it. The idea of moving a digital model closer and closer to the real thing is in fact a basic rationale for building computer models. Of course, there are many models that are extreme simplifications of the real thing where their simplicity is reflected in the main reason why the model is constructed in the first place. But one of the quests in city modelling is to merge social and economic processes with the built environment and to link functional and physical processes to socio-economic representations. One attempt to do this in our group is contained in the model being built for the Queen Elizabeth Park in East London where a 3D model which we casually refer to as Virtual London (ViLo) is being linked to real-time data with this data informing the model in real time (Dawkins et al., 2018). This involves beginning to link the high-frequency city with the low frequency, and this is a quest that has only just begun. It is aided by the internet of things and the deployment of large-scale sensing in the city, which involves ourselves as sensors through our smartphone technologies as well as the physical sensing of energy, traffic, email and related physical and ethereal flows. How close we can get to the real thing is a collective exploration we are about to embark upon as a society. There are enough speculations anyway that we are already living in a ‘simulation’ (Bostrom, 2003), and progress towards digital twins will surely enable us to gain ever deeper insights into the nature of reality and its virtual form. Models as digital twins may be a topical idea whose time will pass but it is all part of the wider scientific effort to make sense of a world where it is clear that the physical and natural are part and parcel of the social and economic: a challenge for integration and multidisciplinarity that characterizes our quest for a science of cities that represents one of the main missions of this journal.