Digital Twins in City Planning

Mike Batty & Jon Reades

Analogue Dreams

Into the Third Dimension

Ask 10 academics and get 10 different answers, but at it’s most basic we’re talking about a virtual representation of a physical object. The term seems to have emerged around 2005, but it’s not hard to trace it back substantially further to, for example, Gelernter (1993)’s Mirror World, or the Day Software Puts the Universe in a Shoebox…. In a built environment context, as the two images on the left show we might even trace it back to the 1980s — top is a model of an English village, and below that is an output from a SOM Architect generated model for 9 US cities — and attempts to understand the relationship between different forms and masses in the real world. On the right is an early attempt to superimpose other contextual information — pollution in this case — on to a 3D surface.

Data-fied Urban Environments

Growing Complexity

Exascale Computation

Something about how exascale and distributed computing allows us to model interactions at unprecedented scale? Also modelling the 2nd and 3rd order interactions which are common in cities but add exponentially to computational complexity? And adding dynamic element to previously static interaction models?

Models of Models

In short, can we embed models into other models such that the output of one is the input to another. Or the same models are running at multiple scales or in multiple contexts? For some kinds of problems this might be relatively straightforward, but in cities — each with its unique history of technologies, policy and political contexts, and data environment — this is non-trivial.

Models and Abstractions

A model is an abstraction or simplification of the real thing, where the model-builder throws away most of the real thing leaving only its essence.

Models can represent the actual system, or they can be a conceptual or mathematical version of the system. However, it’s fairly uncontroversial to acknowledge that they are a simplification of the real thing where choices are made to further the function of the model. Obviously, as our access to ‘compute’ has expanded so has our ambition for the scale of system to be monitored and controlled, but the fundamental limitation remains: our understanding of function will vary from person to person, or group to group, so it’s highly likely that, even if we’re trying to build the same model we’ll end up with something different.

Urban Analytics		High Frequency
Urban Theory1		Low Frequency

The instrumentation of infrastructures and mobility platforms, and the spread of computation into every thing around us, also presents us with the challenge of frequency. For high frequency, you can think of ‘smart’ systems, whether they are the ‘smart trees’ at this year’s Chelsea Garden Show or the ‘smart cities’ that have been a developmental priority in many countries. For most models and twins it’s relatively straightforward to peer a short way into the future.

That’s why, in the high-frequency world of urban analytics it seems plausible that we can use the twin to influence the real thing. Real-time data about the Tube system doesn’t just mirror what’s happening in the Tube, it can be used to shape and change the deployment of people and infrastructures.

But when we try to use high frequency data to forecast what cities might look like in ten, twenty, or thirty years we run into huge issues: indeed, “our current predictions for many social processes are at present not better than two-to-three-day weather forecasting.” So there is a real danger in attempting to use high-frequency data or urban analytics techniques on low frequency problems, and a big part of that is the absence of guardrails when you don’t begin with any kind of theory about the system you are modelling.

1. Or politics!

Shallow copy		Copy by reference
Deep copy		Copy by value

If we pull the ideas from the preceding slides together, there’s a nice analogy to how programmers distinguish between shallow and deep copies of objects and data structures. They mean this in a very particular way about copying references (pointers to the real data) and copying values (the data itself).

Copying by reference is fast and efficient – especially when you’re dealing with large volumes of data – but changes to the reference update the original. Copying by value replicates the object itself in a new memory space where changes to the copy don’t affect the original. So that’s slower, less efficient, but can also be safer. When you copy by value on large data sets you often have to think rather carefully about what data you actually need.

So at the risk of stretching this analogy to breaking point, all of this throws up a deeper meaning as well: perhaps our digital twins copy normative values as well? Like the model-builder, in building a digital twin we make choices about what to keep and what to throw away. Clearly, we also seek to interact with the real-world, but we need to decide if our aim is high- or low-frequency interaction.

If the interaction is low frequency then things get harder: is the flow of information direct, or is there a human-in-the-loop which necessarily going to profoundly alter the nature of the interaction. What we’re getting to here is that maybe there is no such thing as a shallow copy of a city?

The Human in the Loop

\(\ldots\) in social systems, we can define many types of twins that reflect a multitude of theoretical perspectives, each of which reflect different features of the real system

In the absence of deep copy — which we’d argue is impossible due to computational complexity — we need to look closely at what we model and acknowledge their essential incompleteness. So rather than thinking about ‘the one true twin’, we have a plethora of perspectives that are rooted in the slow cycles of theory. They are hierarchical and contradictory, overlapping and multi-scalar. This highlights the importance of human judgement and the need for planners and government to think about what is ‘copied’ into the twin and how its outputs are copied back into the real world…

Without the human-in-the-loop, AI-optimises-everything leaves a system that is higly optimised but almost certainly extremely brittle in the event of catastrophic failures. Efficiency is, paradoxically, an economic problem: the hyper-efficient allocation of ‘resources’ does not lead to the kinds of innovation that arise in places where ‘friction’ and ‘redundancy’ and ‘collision’ create new pathways for development. And our point is that conceptualising the value of the human in all of this is absolutely critical: without that feedback loop it’s highly likely we end up with cities that boring, fragile, and deeply unpleasant.

Thank you!

Arribas-Bel, D., and J. Reades. 2018. “Geography and Computers: Past, Present, and Future.” Geography Compass 12 (e12403). https://doi.org/10.1111/gec3.12403.

Batty, M. 1990. “Invisible Cities.” Environment and Planning B 17 (2):127–30.

———. 2013. “Big Data, Smart Cities and City Planning.” Dialogues in Human Geography 3 (3). SAGE Publications:274–79.

———. 2018. “Digital Twins.” Environment and Planning B: Urban Analytics and City Science. Sage Publications Sage UK: London, England.

———. 2019. “A Map Is Not the Territory, or Is It?” Environment and Planning B: Urban Analytics and City Science. SAGE Publications Sage UK: London, England.

———. 2021. “Multiple Models.” Environment and Planning B: Urban Analytics and City Science. SAGE Publications Sage UK: London, England.

———. 2024. “Digital Twins in City Planning.” Nature Computational Science 4 (3). Nature Publishing Group US New York:192–99.

Batty, M., and R. Milton. 2023. “Building a Digital Twin for British Cities.” arXiv Preprint arXiv:2312.03674.

Gelernter, David. 1993. Mirror Worlds: Or the Day Software Puts the Universe in a Shoebox... How It Will Happen and What It Will Mean. Oxford University Press.

Reades, J., F. Calabrese, A. Sevtsuk, and C. Ratti. 2007. “Cellular census: Explorations in urban data collection.” IEEE Pervasive Computing 6 (3):30–38.

Reades, J., J. De Souza, and P. Hubbard. 2019. “Understanding Urban Gentrification Through Machine Learning.” Urban Studies 56 (5):922–42.