Thesis Review is a collection of conversations, statements and inquiries into the current state of thesis in academia. Thesis projects give a glimpse into the current state of the academic arena while painting a picture for the future of practice. 

Each feature will present a contemporary thesis project through the voice of those that constructed it. This week, we talk to Stanislas Chaillou about his thesis, Bias & Architectural Style: A New Frontier for AI In Architecture, recently presented at the Harvard Graduate School of Design, in the Master of Architecture program. 

What was the intention of your thesis? What aspect(s) of the field of architecture/design did you want to challenge?

The initial intention for this thesis was to structure a consistent framework for what is broadly called “Generative Design”. This field brings concepts found in computation to Architecture, to create potential design options, in a somewhat-autonomous way. My intuition was that Generative Design could find new answers to its shortcomings using Artificial Intelligence. I posit that Generative Design can be understood as a twofold process, which can generate a wide variety of relevant design options while ranking & ordering them according to user-defined criteria and metrics.

In brief, I have prepared four different models trained on different styles (Baroque, Manhattan Unit, Suburban Victorian and Row-House), and studied the behavior of each model, by observing its generated result. I then turned to a challenging site in Manhattan, where the geometry of the plot would normally not allow for a standard development and did my best to fit a housing project on that parcel. For each apartment unit of the development, I used the GAN-model – and therefore, the style- most suited to the constraints. The entire massing was in-filled using my GAN-models, resulting in internal structures & furnishing arrangement varying from one style to the next, from one unit to its neighbor.

Beyond the simple process of generating apartment unit layouts, the aim of my work is to study architectural style learning across a vast array of results and push this concept to its limit.  

How did your thesis change over the course of a year? 

At first, my interest revolved around the strict internal organization of apartment units. My initial attempt was to train some GAN-models to replicate and even create space layouts. The initial results already proved to be quite interesting, as my models were able to recreate some amount of architectural intuition. However, I quickly decided to shift my focus and study a deeper bias, visibly present in those results. 

In fact, since my GAN-models had learned space layout by looking at a specific database of floorplans, their behavior was heavily influenced by forms and typical space organizations present in the initial database. In clear, the architectural style present in the database was being replicated by my GAN-models. 

This is where my thesis pivoted, to tackle architectural style learning. Instead of trying to disentangle style and organization, to create an agnostic set of GAN-models producing “generic” floorplans, I redirected my thesis towards studying the function of style. This notion, coined by Farshid Moussavi in her book The Function of Style asserts that architectural styles carry, beyond their cultural significance, an implicit set of functional rules. These rules drive the composition of floorplans and can be partially captured by my GAN-models.

Do you believe the method of shape generation you developed can also be used as an aid for studying architectural history?

If we think of styles as being the result of architectural history, then yes, I do. And this is maybe where my work could open new possibilities. If there is within each style a deeper set of functional rules, then studying architectural history could potentially be about understanding the evolution over time of these implicit rules. Being able to encapsulate each style could potentially allow us to go beyond the study of precedents, and complement it by unpacking the behavior of GAN-models such as the ones I trained. 

The most tangible example of this idea is what I did early in my thesis: once I had four models trained on each specific style (Baroque, Manhattan Unit, Suburban Victorian and Row-House) I could provide each model with the same set of constraints (same apartment unit footprint & fenestration) and observe how each style would organize space. And of course, for similar constraints, each style came up with its own specific internal structure & logic. Depth, compactness, façade orientation & shape, etc.… are characteristics of a space that are handled very differently by distinct architectural styles.

What was the inspiration for your thesis? Did you venture into other departments at Harvard to gain inspiration?

The inspiration for my thesis came from outside of Architecture: I had the opportunity in 2018 to take a Machine Learning class at MIT. I had studied Computer Science all along my Master in Architecture at Harvard and worked for tech companies, but the encounter with Generative Adversarial Neural Networks (GANs) during a talk given in class caught my attention. 

How do you hope to develop your thesis in practice?

Looking forward, I am less interested in the autonomy of AI or in its ability to converge on objective generative criterion than its ability to encapsulate & emulate some of the unspoken rules of architecture. This “quality with no name” embedded in buildings that Christopher Alexander defines in his book The Timeless Way of Building strikes me as often being the hinge around which revolves the discussion in our discipline. AI is simply a new way to study it. 

In other words, if we can think of floor plans first as compositions, before being strictly the product of engineering, then studying the driving forces of the composition is maybe where AI can offer us some meaningful answers. Concretely, I want to apply my research to create new tools for architects, and new investigative methodologies to study Architecture. On a more theoretical level, I want to keep writing articles, and refining the knowledge base around AI’s inception in our discipline. 

There is a lot to be said, and to be clarified on this topic, and, along with other actors in this field, I hope to help structure this new page of our discipline.