Machine Learning and Architecture

Think Piece about Artificial Intelligence by our valued client, Christopher Beech


As someone who very much enjoys the architectural design process, an interesting question for me is: how can so-called ‘artificial intelligence’ help? To be clear, what I’m not advocating is replacing architects, rather the idea is to use machine learning (that is, pattern recognition) algorithms to broaden the possible range of solutions that would otherwise be too time-consuming to consider. For the sake of simplicity, I’m going to focus on housing.

When an architect or someone with sufficient knowledge looks at a potential site he/she can very quickly form an idea of massing and perhaps even some of the internal configuration. For example, my knowledge has been built up from architectural typology books.

Typology Books


Would it be possible to use such resources to train machine learning algorithms? I don’t see why not, although the technical skills required are beyond my own capabilities. Clearly the breadth and depth of what machine learning could be trained with is beyond the capacity of a human brain. It does raise an interesting copyright question though: if your architectural plans, say of an apartment block, are being used to train a machine learning system, what are the Intellectual Property implications? How does this differ to how an architect already works?

I can envisage machine learning being used to design apartment blocks from the ground up, trying different combinations of apartments to try and understand what exactly could fit on a site. In fact this is exactly how I approached my own project, albeit in a rather more analogue way!


Creativity is nothing without constraints and it is easy to list a number of them:

Alternative massing models and internal configurations could be generated from which the architect could use as the basis for further development. Architectural judgement is still necessary.

It would be wonderful if such a tool was also able to automatically generate the required sunlight and skylight reports, showing the impact on neighbouring buildings. I find it quite frustrating that the design process — iterative by its very nature — is slowed down by having to wait to see if a given massing has too great an impact on its environment. Of course you’d have to tune the algorithms so that every building didn’t end up looking like the LSE Student Centre or New York’s setback architecture.

Which brings me to the next possibility. It would also be possible to train machine learning on elevations. For example if you wanted a classical idiom the input could be Bath and Edinburgh New Town. If you wanted variations on a townhouse theme then the input could be Amsterdam’s canals or their modern equivalent at Borneo Eiland. (Amsterdam’s Eastern Docklands and other docks around the world being exceptions to the general case that the supply of land is fixed.)

Alternatively, you could specify your own facade design and have machine learning calculate the best interior layout. The more complex the problem, say a large apartment block, the more value machine learning could add in optimising the design.

Another data-driven possibility is to use machine learning to recognise Amsterdam’s various architectural features: gables, cornices, overall style, width, number of windows, proportions, etc. and cross-reference that to the register of construction date. This could be done using images from Google Streetview and insights into how styles changed over time could be generated.

Amsterdam buildings by age from


Machine learning algorithms recognise patterns. Hence what I am actually proposing is nothing more than using the latest technology to create modern day architectural pattern books. Just as architects have nothing to fear from pattern books, they have nothing to fear from machine learning.


Christopher Beech is a long-standing client of InsideOut, working with us on two different multi-unit housing schemes, Passivhaus Apartments – Camden and Kilburn Apartments – Camden.