AI takes on city design

Architects and engineers have long harnessed the power of software to automate aspects of the design of buildings and structures, but a new breed of generative design tools based on artificial intelligence is pushing the boundaries of scale, complexity and computational power like never before.

Software products like Delve, by Google’s sister company Sidewalk Labs, and Spacemaker, recently acquired by Autodesk, churn through reams of numerical and contextual data and, based on project priorities, spit out a series of optimised designs for developments, neighbourhoods or even entire city districts.

AI and machine-learning algorithms enable the tools to crunch more data and weigh up a plethora of project considerations much faster than human design teams can achieve.

This can take the stress out of complex urban design work, enabling city planners to explore a greater variety of options than normally possible and leading to better outcomes for things like traffic flows, access to amenities, access to daylight, and pavement space.
In one pioneering recent example, municipal planners for the city of Sofia, in Bulgaria, used Delve to test out a new approach to sustainable development, using geospatial data to analyse an entire ‘city unit’ based on the specific needs of the city and its citizens.

Yet greater reliance on AI also raises concerns over the expanding role of computers in design decision-making. Does AI merely provide a helping hand in the process or wrest control away from the people we entrust to make crucial decisions about the places where we live and work? Will it lead to better quality and more inclusive urban development, or create a generation of cookie-cutter-style buildings and neighbourhoods devoid of character and individuality?

Looking to the future, will we always need a ‘human in the loop’ to vet and approve suggestions made by machines or will technological advancements give them the power to automate the entire process, including decisions on difficult trade-offs?

Dr Dragana Nikolic, lecturer in digital architecture at the University of Reading, says that “when it comes to clear performance criteria, like better energy performance, better transportation systems, better services etc, AI’s superfast analysis of different options and ability to identify optimum solutions is really helpful and quite promising. But when it comes to ‘softer’ areas around social sustainability, community, justice, and ethics etc, there’s still a lot to learn about how AI can address those. This is where we can start having a debate about whether the technology can handle enough variables to make value judgments.”

In a similar way to how organisms evolve in the natural world, generative design exploits software algorithms to automatically produce optimum forms for products and buildings. Input various interdependent parameters and the computer explores all the possible permutations of a solution, generating designs that can be used or become a springboard for new creations.

Recent advances in cloud computation and AI/machine learning have supercharged this process, opening up new opportunities to work with more complex data sets and tackle projects on a much bigger scale.

Spacemaker focuses on the planning phase of building development, where “there are a lot of decisions being made about where to build, how to build and what to build,” explains Carl Christensen, chief technology officer at the firm. “The aim is to remove the errors that can creep in early on before they are locked in for the project’s duration.”

Users, typically architects or urban planners, input data on various physical characteristics, such as terrain and buildings, weather, noise and light conditions, plus local regulations and other preferences. They then ‘weight’ the importance of different parameters, such as the number of rooms that fit within a plot, the amount of sunlight buildings require, and the maximum noise allowed from passing traffic. From all this information the system generates a set of site proposals accompanied by detailed analysis.

It’s possible to explore scenarios at urban scale, at site level or for an individual building. Certain areas of a generated proposal can be locked down if they are successful, while other areas are reprocessed, or some inputs updated. For example, the daylight requirement on just one area may be altered.

Clients, which include Skanska, French property developer Bouygues Immobilier, and Norway’s largest housing developer OBOS, have used the system to identify options for higher-density developments with higher-quality living units, which can maximise the opportunity for return on investment. For example, if apartments have more access to daylight and better views, they may be easier to sell, improving the wellbeing of occupants and gaining the support of local authorities.

‘Design better cities, faster, with less risk’ is the aspirational claim made by Delve on its website. The firm applies a similar approach to Spacemaker in order to scope out a range of urban developments from inception stage through to a fully permitted masterplan.

The cloud-based software grapples with competing project considerations for things like density, daylight, amenity access and infrastructure, also taking into account the project constraints and site context.

The highest-performing options, from a list of several thousand, are automatically identified and ranked based on how well they meet a set of pre-defined project priorities. A financial model is also provided for individual segments of every design option.

A close physical match to reality is achieved using integrated ‘generators’ that mimic actual city environments. For example, a building generator can produce a broad set of building types that are designed to be buildable, attractive and fit in with the local context.

Rather than simply crank out a batch of districts that all look homogenous, the best-scoring proposals are selected to differ from one another in strategy, using an AI-powered process that “mimics human aesthetic insight”.

Douwe Osinga, director and software engineer at Sidewalk Labs, explains: “We train the machine-learning model, using computer-vision models, to cluster districts based on how much they look alike. If two models look alike it’s not all that productive to serve them both to the human. We want districts that score well but look different.”

‘Our tool supercharges intuition with data and smart analytics, but it is people that turn this information into strategic approaches, that create long-term and sustainable solutions for our cities.’

In the high-stakes world of urban property development and master planning, AI-based generative design tools can lead to valuable insights and productivity gains.

The city of Sofia’s planning authority, Sofiaplan, used Delve to test out the viability of hypothetical new-build developments, both in terms of construction impacts and impacts on existing social infrastructure, including current and forecast population, engineering systems, the availability of public and private green spaces, and noise pollution.

Darina Manolova, head of parametric planning at Sofiaplan, says: “Delve simplifies the preliminary study process and gives a broader perspective on the possibilities of the project, many of which might have not been considered at all. Implementing tools like this into city planning processes could become an important step for city planners and decision-makers to have a better understanding of the present state of the environment and assess the necessary steps for balanced growth of their cities.”

A prominent Japanese real-estate developer recently used Delve to test design scenarios for a major mixed-use project in a major city, with a focus on optimisation for height limits, financial performance and quality of life.

The tool recommended that the two largest uses, office and hotel, were reduced by 10 per cent and 12 per cent respectively, with an additional 40,000 square metres of new residential space. A view-scoring matrix helped boost hotel and office views to existing parks and the water by over 40 per cent. Together, this evidence encouraged the client to underwrite 5 per cent higher rents than originally modelled.

Furthermore, a test exercise in the UK saw developer Quintain employ Delve to optimise a development of build-to-rent homes at a large site in north London. Although the tool was ultimately unable to improve on options previously generated using traditional means, one high-performing variant added nearly 200 units, improved daylight access, and expanded open space by 11 per cent.

French architects Valode & Pistre tested Spacemaker internally to measure productivity gains when creating proposals for an 80,000 square metre office and residential development, compared with the traditional way.

It took just 42 hours to design the first concept using Spacemaker (an initial proposal from the software was ready after only 34 hours), versus 56 hours, or seven working days, designing the same concept conventionally.

This freed architects to focus on more creative tasks, says Annalisa De Maestri, director of BIM at Valode & Pistre: “One of our big challenges when preparing concepts is that architects typically spend three days a week modelling in software and just two days thinking about a project. Spacemaker showed us that we can invest more of our time in the creative process.”

Greater reliance on computer-enabled decision making in urban planning may boost efficiencies, but it is controversial given the potential long-term social, economic and cultural impacts once development is completed.

Both Delve and Spacemaker emphasise the role of their products as an ‘enabler’, and the importance of pairing AI with human intuition, which is necessary to understand local nuances, such as aesthetic preferences, political environments, zoning requirements and the web of complex relationships inherent in urban planning.

As things stand, human input remains critical, says Julian Tollast, head of masterplanning and design at Quintain: “[This type of AI/generative technology] seems less successful for qualitative criteria and that is where the ‘human process’ is more successful... It could certainly carry out the numerical aspects of design generation and evaluation but doesn’t appear to cope well with aesthetics or the softer criteria.”

Carl Christensen, CTO and co-founder of Spacemaker, adds: “Our tool supercharges intuition with data and smart analytics, but it is people that turn this information into strategic approaches, that create long-term and sustainable solutions for our cities.”

Nevertheless, it’s not hard to imagine the dynamic between human and machine changing in future as the technology becomes smarter and able to handle more complex data and decision making.

More oversight may be required to ensure that conflicting priorities are properly balanced, like public wellbeing, and access to greenery and open space is prioritised over economic benefits linked to business development and more roads.

In this context, the integrity of data being fed into tools could become increasingly important. “As with any computer tool, if bad data goes in, bad data comes out,” says Dr Nikolic at the University of Reading. “Data can be highly biased and derive from a number of different stakeholders and sources. It could relate to the economic status of the residents, or other contextual factors within a location. The more data that becomes available, and the more we ensure it is fair and unbiased, the more useful it will be in informing our decisions.”

More sophisticated number-crunching could give a creative boost to the design process. Algorithms could act as a powerful ‘creativity pump’, used to generate innovative and unexpected solutions that human teams use as a jump-off point to create exciting new architecture and spaces: “Like fashion shows where big designers roll out impossible-to-wear designs that nobody actually thinks regular people are going to wear, but high street brands look to them for ideas,” says Christoph Salge, lecturer at the School of Computer Science at the University of Hertfordshire and organiser of the Generative Design in Minecraft Settlement Generation Competition.

AI/generative design tools could also improve opportunities for citizen engagement in design, by being able to analyse and weigh up the conflicting needs and wishes of many more participants than current processes permit.

“There’s a big urge nowadays to integrate different perspectives into city planning,” says Salge. “Maybe a procedural or generative design approach could take in more views and be more sensitive towards them than even a human could be.”

Advances in technology often outpace humans’ ability to understand their consequences. The more we observe and understand the complex technical and ethical issues presented by AI and generative design, the more likely it is that the cities they aim to shape will reflect the needs of the people that live and work within them.