Our work finds itself at the intersection of Architecture and Artificial Intelligence. The former is the topic, the latter the method. Building upon preexisting research in this field, this project organizes the encounter of Architecture & AI, investigates the resulting architectural forms that it produces, through building and masterplan design.
Architecture is here understood as the intersection between Style and Organization. On one hand, we consider buildings as vectors of a cultural significance, that express through their geometry, taxonomy, typology, and decoration a certain style. Baroque, Roman, Gothic, Modern, Contemporary: as many architectural styles that can be found through a careful study of floor plans. On the other hand, buildings are the product of engineering and science, answering to strict frameworks and rules -building codes, ergonomics, energetic efficiency, egress, program, etc — that can be found as we read a floor plan. This organizational imperative completes our definition of Architecture and drives our investigation.
Artificial Intelligence is here employed, using one of its most recent fields of investigation: Generative Adversarial Networks.If creating standard apartments can be achieved using such a technic, pushing the boundaries of our models motivated our work. AI & GANs can, in fact, offer quite remarkable flexibility to tackle seemingly highly constrained problems. In the case of floor plans layout, we studied their ability to offer surprising architectural solutions, while being challenged to deal with complex shapes (unusual building footprints, or program repartition).
Our work tackles the dual imperative of style and organization, by offering AI systems trying to emulate given architectural styles, and spatial organizations. First, we developed the idea of style transfer, where a given floor plan is ripped off of its wall thickness(A) and dressed back with a new wall stylistic(B).
Then, we offer a collection of building plans and master plans, where our models try to recreate a meaningful internal organization of given floorplans. By feeding our models with shapes, taxonomies, and typologies unusual to architectural design, our models are challenged to find each building’s program repartition (marked with colors), fenestration and furnishing.
The variety and richness of created floorplans are striking. As opposed to the legibility of parametric design, AI seems to offer a fascinating breadth of options and an intrinsic complexity. Beyond this experimental investigation, we hope to open the door to new ways to conceive design as an interaction between humans and machines.
What prompted the project?
This project stems from the intersection between Artificial Intelligence’s recent developments, and Parametricism’s shortcomings.
AI, and its subfield Neural Networks, have seen quite fascinating evolutions over the past 4 years, bringing more manageable and “intelligent” tools, able to generate patterns, geometries, and images. At the same time, Parametricism in its current form does not allow sufficient flexibility to designers, while missing the ability to grasp high-level constraints. Not everything in Architecture pertains strictly to efficiency, and declaring variables –or parameters- is often an over-simplification of design, leading to predictable & almost-templated generated floorplans.
AI, and more specifically Generative Adversarial Neural Networks (GANs), flips this logic: instead of asking for an explicit encoding of parameters, it lets the machine “learn” patterns found among examples (in the “training set” as we call it). In return, the algorithm will know how to emulate the characteristics found among the observed data.
Following this logic I have trained an array of GAN-models to learn different steps of the design process of apartment floor plans: one model learned how to draw a building footprint, the next one knows how to layout rooms across the previously generated footprint, while the last model turns each room from the previous step into a fully furnished interior. I have also trained an array of other models to learn architectural styles, and their internal logic. Using all these AI models, I finally have designed entire buildings and master plans.
What questions does the project raise and which does it answer?
This project raises the question of our usage of technology as architects. It offers a new way forward, where we could potentially use AI to better encode space’s logic, and go beyond Parametricism’s shortcomings. In clear, my belief is that a statistical approach to design conception shapes AI’s potential for Architecture. Its less-deterministic and more-holistic character is undoubtedly a chance for our field. Rather than using machines to optimize a set of variables, relying on them to extract significant qualities and mimicking them all along the design process is a paradigm shift.
In that regard, this project also tries to answer what would the architect’s place be if such tools would be developed further. The answer is that we, architects, have the opportunity to shape this new toolset. AI is only as “clever” as our ability to teach it what is relevant in Design & Architecture. As I was training and tuning my GAN-models, it became very clear that my architectural knowledge was, in fact, the skillset I used most to be able to obtain well-functioning tools.
What is your take on parametric tools as Finch? What are for you the strengths and weaknesses of technologies like these?
Finch has found a sweet spot in the world of Parametricism. Such tools are called “procedural”, as they follow a clear set of user-defined rules, like a procedure. This type of algorithms can quickly sketch out reasonable design options. However, they are deeply anchored into the logics of Parametricism: the set of parameters driving the design is set, the rules are fixed, and the results therefore limited to certain use cases. An algorithm drawing units for a single-loaded corridor typology will have a hard time tackling another type of typology. In other words, there is here a lack of flexibility that, ideally, a design tool should allow for.
This is where AI comes in: there is a level of ubiquity and flexibility that is quite fascinating in AI’s ability to understand and replicate layouts & space’s structures. An AI-model trained on seemingly standard floorplans will be able –to some extent– to tackle in a strikingly relevant way new sets of constraints while yielding relevant design options. This flexibility is illustrated by the GIFs in my project: as the footprint of the apartment unit changes, taking seemingly impossible shapes, the algorithm is still able to find a reasonable way to layout rooms and walls across space. The result might not be perfect yet, as this approach is still very much experimental, but the overall layout of rooms and elements in space is often reasonable.
To what extent could this new methodology affect the role and possibilities of the architect?
Generative Design, or simply put, the utilization of software and algorithms to offer potential design options adds an additional layer to the brainstorming happening is any architectural project. In its simplest version -Parametricism- it offers architects a way to resolve constraints, in its most interesting form -AI-, it becomes a source of inspiration, providing us with an almost-limitless pool of ideas.
To answer the question of the architect’s role, I want to touch upon a concept defined by Prof. Andrew Witt: the idea of “Grey Boxing”. This expression sums up perfectly the fact that most AI models, and algorithms coming from computer science do not “ship” as airtight technologies, perfectly packaged into a software. As opposed to BIM software or 3D modeling software we are used to as architects, the hacky nature of computer science forces the user to shape his/her toolset. As opposed to the “black box” that some software sometimes are, a grey box allows a partial control over the structure and functionality of the tool. For my thesis, before being able to design any buildings, I had to create my own set of tools. If Grasshopper lifts a bit the veil, and gives us a taste of the Grey Boxing idea, using some tools found in computer science requires a more experimental attitude. That is in itself a redefinition of our role as architects: we have the opportunity here to encode our vision/style/ideas directly in our tools. As we will train AI-algorithms to assist us, we will have the chance to educate them to our own sensibility.
When talking about the digitalisation of the discipline when would you define this? Could we talk about a hyper-digitalisation?
I think the notion of “digitalisation” – and “hyper-digitalisation” even- too often refers to the on-going superficial change in our industry. It is simply referring to a transition happening at every scale: formats, processes, even job titles. From CAD to BIM, from drawings to models, from “Architect” to “BIM specialist”.
To this quite shallow manifestation of “digitalisation”, our interest as architects should lie in understanding and shaping the revolution of the creative process that is concurrently happening. It is deeper and way more fascinating evolution. If it is expressed through the medium of software and programming, its conceptual under-pinings completely encompass the traditional opposition of digital vs analog, or the strict constraints imposed by computer science or any form of engineering. To answer “How should we teach an architectural AI, and what should we teach it?”, we open up a profound discussion about Architecture’s purpose. In reverse, answering this question forces us to come up with definitions and a structured representation of our discipline, to teach the machine. This effort, this need for verbalization is a meaningful exercise, at a time where our discipline is struggling between confusion & self-doubt.
What is for you the architect's most important tool?
Creativity. Shear creativity. I think the toolset is only as relevant as our ability, as architects, to drive the creative process with talent and purpose. To a creative mind, AI only is an opportunity to push design further. It is not a substitute for creativity, but a catalyzer.
However, as compared to previous tools available to designers, AI can grasp higher levels of abstraction, such as style, space composition & syntax, and others. Also, it is a new field of investigation. Beyond the practice, and going back to architectural research, AI can shed new light on common research topics. Notions like precedents, style, bias, efficiency find here a renewed methodology and analytical framework. My thesis at the GSD investigated the bias induced by architectural style on space and its internal structures. Many other studies, taking the same approach, could clear up some existing blind spots of our discipline.