Atoll is an urban design and governance platform driven by deep learning.

Through simulating the interplay of multiple urban simulations it creates a space for new forms of discontiguous urbanism.


Atoll is an urban design and governance platform driven by deep learning, informing the spatial planning of cities and creating forms of discontigous urbanism.

The platform consists of 3 main components. The MIM and ARRAY which make up the Atoll simulation space, and a series of planning tools aimed at urban planners: INSTANCE, SNIPPET and ECHO.

And CONDENCER – a participatory platform which allows communities and individuals to create their own models for consideration within urban simulation.


MIM - Russian for Mimic, is the main input for the Atoll platform. An evolution of building and city information management systems, MIM is a simulation management system that creates and manages profiles of different urban simulations.

The use of simulation in all aspects of urban development is creating a new paradigm, which places the collection of data and the modelling of systems at centre stage. However existing definitions and categorisations of urban indicators fail to address the complex interplays and temporalities of urban events.

The Atoll platform takes multiple data types. Rather than being restricted by needing tidy clean data MIM is able to absorb traditional data models from BIM, GIS , Digital Terrain Models and other static models of the city and also dynamic data from other platforms and sensors, which can be added too as new networks are added in the city.


By training ARRAY on the profiles of different clusters within the city it is able to create new clusters that mimic existing aspects of the city. Training these simulations becomes the way in which you make the city.

Generative adversarial networks, through the use of two neural networks contesting one other in a zero-sum game framework (thus “adversarial”) can learn to mimic various distributions of data and are therefore valuable in generating test dataset’s when these are not readily available.

Once these profiles have been synthesised by Array, the reverse process can take place, with the data being decompressed from profiles to create entirely new city models which retain the characteristics, behaviours and functions of the original, simulating environments, flows and and new urban configurations. These simulated spaces can then be used within 
a second layer of simulation.


ARRAY is based on the architecture of GAN’s - generative adversarial networks - a form of deep neural net which pits two nets against each other in order to generate new things that have never existed.

One way to think about generative algorithms is that they do the opposite to discriminative algorithms. Instead of predicting a label given certain features, they attempt to predict features given a certain label. In the model of a GAN, one neural network, called the generator, generates new data instances, while the other, the discriminator, evaluates them for authenticity; the discriminator decides whether each instance of data it reviews belongs to an actual dataset or not.


INSTANCE is the first planning tool, that simulates modifications on decoded space propagating through various profiles to show the impacts on the urban fabric and highlight other unconsidered changes.


SNIPPET is based on the infill function of Generative Adversarial Networks. By synthesising new configurations of urban space from multiple different models, either from the same city or a different one, SNIPPET can simulate the affects of an infinite number of alternatives for selected urban regions.


ECHO acts as a training space for AI driven services. Allowing an infinite number of soft, oscillating landscapes to play out over various time-scales.


As the Atoll platform evolves, multiple models of the city begin to exist at the same time, as fields of correlations are created. These models of clusters of the city can each be seen as part of a larger competing network analogous to the nesting of multiple Generative Adversarial Networks.

Each taking on the role of a generator within the GAN model, passing a city level discriminator models of populations, services and future planning, zoning and budgetary implementations. defining its needs in order to shift the manner in which it is classified.


These models however are dependent on the data that they are able to capture, from residents, visitors, networks of sensors or transport systems. In current smart city imaginaries we are implicitly part of models generated for simulation, as our clicks, swipes, and data exhaust is cross-referenced.

The Atoll platform allows anyone to define a model for consideration, creating a space for community engagement and allowing a form of translation for the informal shadow systems and myriad of local intelligences and situated knowledges already existing in a cities material assemblages and social assemblies.

Atoll is going live [soon]


Atoll was developed as part of the New Normal education programme at Strelka Institute for Media, Architecture and Design.

Tom Pearson
Leo Stuckardt
Nataliya Mezhetskaya
Artem Konevskikh

Contact us on