The article doesn’t go into much detail, but it hints at the use of a technique that I’ve suspected might have some promise for building design - finding a way to convert a 2D set of drawings or a 3D model to a text-based representation, then using text-based AI tools like GPT-3 to generate new buildings based on input parameters.
Currently the building design process is time consuming and labor intensive, I think there’s lots of potential for automating large swaths of it.
> finding a way to convert a 2D set of drawings or a 3D model to a text-based representation, then using text-based AI tools like GPT-3 to generate new buildings based on input parameters.
Kind of sounds like shape grammars [1] which can compose grammatical rules into designs and vice versa.
I think it would be interesting to use a shape grammar engine with Reinforcement learning, where the design is a state, action is a shape rule, and reward is some performance metric. The advantage over GANS (which seems to be the most common use of ML in architecture/design circles) is the integration of cumulative reward maximization better captures the way designs are incrementally built up.
This is really interesting, like a computer algebra system[1] but for geometry. Seems like it could be extended to cover the non-shape based documentation (material properties, design loads, etc) as well.
One thing I've never seen is optimization techniques paired up with any sort of in-depth cost data. It's usually based on naïve extrapolations based on total material weight or volume, and doesn't take into account things like crane costs (a function of max component weight and building extents) local material availability, component fixed costs, number of connections, logistics costs, etc. There's a ton of low-hanging fruit in design optimization that you could pick before you even had to reach for AI-based methods, though the AI methods are obviously the long term future of the space.
That sounds like it might make the exercise more tractable. Without some sort of middleware, the article seems to be saying the researchers lack traction.
Or maybe the problem is a lack of data.
What if each building was a dataset that included digital versions of the client brief, the CAD files, the BOM and procurement folio, and a 5 year subjective client review? If there were millions of these datasets for training, the same learning could occur as with the shape grammar approach.
Perhaps the limitation today is not enough data, irrespective of data format.
Well, it seems efforts to introduce AI are going to happen everywhere that human create or look at data. Some uses seem pretty effective - neural nets sorting images of stars seems to have found some interesting things. Some uses seem fairly problematic - modern buildings seem like a complete disaster already to my naive eyes and less attention might not improve things. Moreover, as far as I know, the job of an architect is to make a building that is aesthetically pleasing and balanced but which also satisfies a lot of specific constraints - has a bathroom, doors don't knock against each other when opened, sufficient lighting etc.
The thing is multiple constraint satisfaction part seems to be where neural nets seems to fail. Self-driving cars work mostly yet by not satisfying all constraints, they seem destined to be "always five years out". GPT-III writes paragraphs that seem pretty well written yet it contracts itself quite often 'cause it's duplicating most frequent phrases, not considering the logical constraints humans (sometimes) need to take into account.
Which is to say, I'd imagine AI generated buildings would seem plausible till you noticed the doors to nowhere and similar stuff.
The design of feasible layouts is definitely one of the trickier parts of applying ML to architecture. I know of one masters and one PhD student working in this domain, and it's so hard to achieve human-quality spatial layouts.
However, that is just one part of building design, and some of the other constraints of architecture (to use your language) lend themselves more to ML/AI solutions, and I'm seeing crazy progress in domains like structural or energy optimization. I'm pretty excited by the potential in this space, personally.
I'd love to see architecture driven by tools that show a projection of how a building might look after a few decades of weathering and, well, "realistic" amounts of maintenance.
Too many current architectural aesthetics are relying on the simple mechanism of focusing the eye on how impeccably fresh it's surfaces are. The buildings will then look impressive while they are impeccably fresh, and just rotten afterwards. Because the eye is then still focused on the same surfaces, now showcasing imperfections. "Minimalistic design" architecture rarely ages well unless aging was deliberately made part of the plan. Highly ornamental architecture however often works quite nicely after aging (or even partial decay), it's as if the visible aging gives the ornaments a form of authority they lack while new.
I don't know wether it would be a simulation approach ("where does how much rain water run down from window sills") or if some clever deep learning tricks similar to style transfer could be used or a combination. The "how" doesn't matter for the effect it could have on how architects and their customers think during the design phase.
Marketing a tool like that would not be easy, but with a sustainability angle it should not be impossible either.
8 comments
[ 3.0 ms ] story [ 29.3 ms ] threadCurrently the building design process is time consuming and labor intensive, I think there’s lots of potential for automating large swaths of it.
Kind of sounds like shape grammars [1] which can compose grammatical rules into designs and vice versa.
I think it would be interesting to use a shape grammar engine with Reinforcement learning, where the design is a state, action is a shape rule, and reward is some performance metric. The advantage over GANS (which seems to be the most common use of ML in architecture/design circles) is the integration of cumulative reward maximization better captures the way designs are incrementally built up.
[1] https://en.wikipedia.org/wiki/Shape_grammar
One thing I've never seen is optimization techniques paired up with any sort of in-depth cost data. It's usually based on naïve extrapolations based on total material weight or volume, and doesn't take into account things like crane costs (a function of max component weight and building extents) local material availability, component fixed costs, number of connections, logistics costs, etc. There's a ton of low-hanging fruit in design optimization that you could pick before you even had to reach for AI-based methods, though the AI methods are obviously the long term future of the space.
[1] - http://www.math.wpi.edu/IQP/BVCalcHist/calc5.html#:~:text=A%....
That sounds like it might make the exercise more tractable. Without some sort of middleware, the article seems to be saying the researchers lack traction.
Or maybe the problem is a lack of data.
What if each building was a dataset that included digital versions of the client brief, the CAD files, the BOM and procurement folio, and a 5 year subjective client review? If there were millions of these datasets for training, the same learning could occur as with the shape grammar approach.
Perhaps the limitation today is not enough data, irrespective of data format.
The thing is multiple constraint satisfaction part seems to be where neural nets seems to fail. Self-driving cars work mostly yet by not satisfying all constraints, they seem destined to be "always five years out". GPT-III writes paragraphs that seem pretty well written yet it contracts itself quite often 'cause it's duplicating most frequent phrases, not considering the logical constraints humans (sometimes) need to take into account.
Which is to say, I'd imagine AI generated buildings would seem plausible till you noticed the doors to nowhere and similar stuff.
However, that is just one part of building design, and some of the other constraints of architecture (to use your language) lend themselves more to ML/AI solutions, and I'm seeing crazy progress in domains like structural or energy optimization. I'm pretty excited by the potential in this space, personally.
Too many current architectural aesthetics are relying on the simple mechanism of focusing the eye on how impeccably fresh it's surfaces are. The buildings will then look impressive while they are impeccably fresh, and just rotten afterwards. Because the eye is then still focused on the same surfaces, now showcasing imperfections. "Minimalistic design" architecture rarely ages well unless aging was deliberately made part of the plan. Highly ornamental architecture however often works quite nicely after aging (or even partial decay), it's as if the visible aging gives the ornaments a form of authority they lack while new.
I don't know wether it would be a simulation approach ("where does how much rain water run down from window sills") or if some clever deep learning tricks similar to style transfer could be used or a combination. The "how" doesn't matter for the effect it could have on how architects and their customers think during the design phase.
Marketing a tool like that would not be easy, but with a sustainability angle it should not be impossible either.