AI + Architecture

Towards a New Approach

Stanislas Chaillou | Harvard GSD | 2019
 Expliquer.
“Explain.”

In 2014, the Ecole Normale Supérieur, one of the most competitive university
in France, gave to its candidates for the entrance exam the following subject:
“Explain”. No subtitle, no annex. An 8-hour exam sat by more than 15.000
candidates, came down to this single word.

Far from the public outcry that immediately followed this event, we see this
injunction as the single most valuable responsibility of our time: explaining.
If one thing is to have convictions, another one is to find the right words to
teach others.
2 3
The age of machines will not abolish this truth. To the contrary, creating
the right taxonomy will more than ever be a crucial skill to configure the
new forms of intelligence expected to leverage Architects’ capabilities. But
it is not a top-down exercise anymore: the machine would not just meekly
execute the instruction transmitted by the “right word”, he may in its turn
suggest its own “optional words”. In fact, the advent of Artificial Intelligence,
or AI, is creating the conditions of a reflexive empowerment: the machine
could become a trustworthy “assistant” provided professionals educate it, or
properly “explain” the job.

Stanislas Chaillou, Harvard GSD, May 2019

© 2019 Stanislas Chaillou, All Rights Reserved.
 Acknowledgments
This thesis is the result of an intense academic journey, halfway between
Switzerland and the US. It is certainly the result of ideas and influences found
in each system; a blend between Swiss rigor and American optimism, with a
dash of French Cartesian thinking. Along the way, it is thanks to inspiring and
open-minded people that I have found the resources to carry on.

I am thankful to Andrew Witt, who advised me all along this thesis; his
positivity and intellectual curiosity have been the ideal springboard. Through
his work, he has inspired at the GSD an entire generation of new practionners,
I am just one of the many.

4 None of this would have ever happened without the unfailing support of my 5
Mum, Dad and sisters: leur apport a été constant et parfait, tant sur un plan
émotionnel qu’intellectuel. Il aurait été impensable d’avoir manqué à mes
études, tant leur présence a rendu ces sept dernières années la source d’un
épanouissement profond.

To the Flux.io Team that catalyzed my intuitions, and trusted me early on, I
am also grateful: Thomas, Anthony, Nicolas. Their advices have shaped the
orientation I chose at Harvard, and this book would not be here today without
them.

Finally, I would like to acknowledge a broader idea, my home country:

France. If I was not fortunate enough to study back home, I am aware of the
intellectual structure, and cultural heritage that this country has given me.
From the singularity of its history, to the breadth of its culture, it is an endless
well of knowledge. Too often taken for granted, and rarely acknowledged for
its richness, I want to end this page by reaffirming how France has offered me
the intellectual anchor & roots that have shaped my academic journey.
 Introduction 9

Table of Content
I. The Advent of Architectural AI 10
A. A Four Period Sequence 11
B. A Continuous Progress 21
C. Closing Remarks 23
D. Framework 25

II. Generate 27

A. Precedents 31
B. Organization 33
Footprint 37
Room Split 41
Furnishing 45
Room Rendering 49
Going Further 51
Organization – Takeaways 61
C. Style 63
Bias & The Emergence of Style 65

6 Four Styles 67
7
Application 83
Style – Takeaways 133

III. Qualify 135

A. Footprint 137
B. Program 141
C. Orientation 149
D. Thickness & Texture 157
E. Connectivity 163
F. Circulation 169

IV. Mapping & Browsing 175

V. Conclusion 183

Bibliography 185

Digital Resources 187

 steps across different scales (urban scale, building
scale, unit scale). Then, in order to harness these
two realities, we have chosen nested Generative
Adversarial Neural Networks or GANs. Such models

Introduction enable us to capture more complexity across

encountered floor plans and to break down the
complexity by tackling problems through successive
steps. Each step corresponding to a given model,
specifically trained for this particular task, the
Artificial Intelligence, as a discipline, has already been process can eventually evidence the possible back
permeating countless fields, bringing means and and forth between humans and machines.
methods to previously unresolved challenges, across
industries. The advent of AI in Architecture is still Plans are indeed a high-dimensional problem, at
in its early days but offers promising results. More the crossroad of quantifiable technics, and more
than a mere opportunity, such potential represents qualitative properties. The study of architectural
8 for us a major step ahead, about to reshape the precedent remains too often a hazardous process, 9
architectural discipline. that negates the richness of the number of existing
resources while lacking in analytical rigor. Our
Our work proposes to evidence this promise when methodology, inspired by current Data Science
applied to the built environment. Specifically, methodologies, aims at qualifying floor plans.
we offer to apply AI to floor plans analysis and Through the creation of 6 metrics, we propose a
generation. Our ultimate goal is three-fold: (1) to framework that captures architecturally relevant
generate floor plans i.e. optimize the generation parameters of floor plans. On one hand, Footprint
of a large and highly diverse quantity of floor plan Shape, Orientation, Thickness & Texture are three
designs, (2) to qualify floor plans i.e. offer a proper metrics capturing the essence of a given floor plan’s
classification methodology (3) to allow users to style. On the other hand, Program, Connectivity,
“browse” through generated design options. and Circulation are meant to depict the essence of
any floor plan organization.
Our methodology follows two main intuitions (1) the
creation of building plans is a non-trivial technical In a nutshell, the machine, once the extension
challenge, although encompassing standard of our pencil, can today be leveraged to map
optimization technics, and (2) the design of space architectural knowledge, and trained to assist us
is a sequential process, requiring successive design in creating viable design options.
 a disruption, we want to see here a continuity that led architectural grammar, contributing to simplify and
Architecture through successive evolutions until today. rationalize building design.
Modularity, Computational Design, Parametricism and Theorized for the Bauhaus by Walter Gropius, as early
finally Artificial Intelligence are to us the four intricated as 1920, the modular grid was carrying the hope of
steps of a slow-paced transition. Beyond the historical technical simplicity and affordability. Coming from
background, we posit that this evolution is the different directions, modularity arose at first as a
wireframe of a radical improvement in architectural topic of investigation for academics and practitioners.

I
conception. Gropius initially introduced the idea of “Baukasten”, a
typical module to be then aggregated through strict
assembly rules. This systematicity will be echoed one
A. A Four Period Sequence year later with Le Corbusier’s “Modulor”. By applying the

The Advent of
modular rigor down to the human scale, Le Corbusier,
as of 1946, offered a holistic implementation of the
Modularity, Computational Design, Parametricism modular principles. The built environment dimensions

Architectural AI
and finally Artificial Intelligence are not air-tight would be aligned on key metrics and ratios derived
steps, independent of one another: each period from the human body. And indeed, from “La Tourette”
interpenetrates and borrows from the precedents. It to the “Unité d’habitation” in Marseille, Le Corbusier
is why, when looking backward at history it is critical
to distinguish two levels of creation: inventions &
10 innovations. Inventions stem from academic research, 11
while innovations are induced by inventions. In
The practice of Architecture, its methods, traditions, architecture, Innovations actually shape a continuously
and know-how are today at the center of passionate moving practice. A practice which has been playing
debates. Challenged by outsiders, arriving with on the back and forth between periods, inventions
new practices, and questioned from within, as & innovations. From there, our chronology aims at
practitioners doubt of its current state, Architecture demonstrating the deeply interwoven evolutions
is undergoing a truly profound (r)evolution. of the computational and the architectural fields Figure 1: Corbusier’s Modulor
before introducing the age of architectural-AI, as a
Among the factors that will leave a lasting impact culminating point. It is why rebuilding the context and systematized the dimensions and spans to match the
on our discipline, technology certainly is one of the the highlights of the recent history of our discipline is a prescription of the “Modulor”. With Buckminster Fueller,
main vectors at play. The inception of technological prerequisite to our work. however, Modularity rapidly evolved towards a more
solutions at every step of the value chain has already integrated vision embedding building systems within
significantly transformed Architecture. The conception Modular Systems the module as exemplified by the Dymaxion House.
of buildings has in fact already started a slow This attempt pushed to its extreme the possibility of
transformation: first by leveraging new construction Modularity could be set as the starting point of modular housing, setting a vibrant precedent, and
technics, then by developing adequate software, and systematic architectural design. Initiated in the early proof-of-concept for the industry.
eventually today by introducing statistical computing 30’s, the advent of modular construction brought Thereafter, following these early theorists, architects
capabilities (including Data Science & AI). Rather than to the conception phase both a language and an were invited to bend their design ethos to the
 imperative of the matrix, and by the same token, to its system of rules, Modularity is remaining today an generation of computer scientists and architects will boundary of assembly systems and allow for new
transfer part of the technicality of building design to underlying constructive principle still vivid throughout create a new field of research: Computational Design. shapes & building geometries. Gehry Technologies,
the logic of the module. Less hassle, less costs, more the practice. The Architecture Machine Group (AMG) at MIT, led by founded by Gehry and Jim Glymph in the 80’s typically
predictability. Modularity would then swiftly extend Professor Nicholas Negroponte is probably its most used early CAD-CAM software — such as CATIA — from
to the industry as a whole: the Winslow Ames House, Computational design Dassault Systems to tackle complex geometric
built by Professor Robert W. McLaughlin in 1933, and problems. Setting here the precedent for 30 years
first large-scale modular project in the world, was At the turn the 80’s, as the complexity of modular of Computational Design, Gehry Technology would
perceived as a major breakthrough, as much as the very systems was soaring, the advent of computation demonstrate the value of computation to architects,
brought back feasibility and scalability to modular provoking a landslide in the profession. Over the next 15
design. Beyond the resurrection of the module, the years, the irresistible growth of computational power
systematicity of rule-based design was somehow & data storage capacities, combined with increasingly
rehabilitated. affordable and more user-friendly machines,
massively facilitated the adoption of 3D-design
Coming from different directions, a high-level software. Architects rapidly endorsed the new system
reflection about the potential of computational design on the base of a clear rationale: Computational Design
started as early as the mid-50’s, within an adjacent (1) allows a rigorous control of geometry, boosting
discipline: Engineering. In 1959, Professor Patrick design’s reliability, feasibility and cost, (2) facilitates and
Hanratty released PRONTO, the first prototype of CAD eases collaboration among designers, (3) and finally
Figure 2: Plugin City by Archigram
12 (Computer Assisted Drawing) software, geared towards enables more design iterations than traditional hand- 13
expressive Habitat 67 from Moshe Safdie. City planning engineered parts design. The possibility offered by sketching could afford. More tests & more options for
even got influenced at the turn of the 60’s, when such software, coupled to the potential computational better resulting designs.
Figure 3: URBAN 5, AMG MIT
projects like the “Plugin City” of Archigram developed power fast-paced evolution, jumped start a discussion
the possibility of modular cities. Through the constant within the architectural field. Soon after, Christopher exemplary embodiment. Negroponte’s book “The However, along the way, as designers were engaging
assemblage and dismantlement of modules, fitted Alexander, architect and then professor at U.C. Architecture Machine” (1970) encapsulates the essence with Computational Design, a couple of shortcomings
on a three-dimensional structural matrix, cities could Berkeley started the discussion by laying down the key of the AMG’s mission: investigating how machines can eventually arose. In particular, the repetitiveness of
find a renewed logic, addressing both the possibility of principles for Computational Design. In his “Notes on enhance the creative process, and more specifically, certain tasks, and the lack of control over complex
growth and -as always- the imperative of feasibility. the Synthesis of Form” (1964) and later, in “a Pattern the architectural production as a whole. Culminating geometric shapes became serious impediments. Those
Language” (1968), Alexander theorized why and how with the release of projects URBAN II and later URBAN paved the way to a brand-new movement which was
However, connecting the grid, the modules, and computers should be used to address the question V, this group will then demonstrate, even before emerging within Computation Design: Parametricism.
the assembly systems through mechanistic rules of shape design. His early understanding of software industry would engage in any effort, the potential of
eventually led to a quasi-gamification of a LEGO-like potential for design was deeply contrasting with the CAD applied to space design.
conception of Architecture. But practice cannot be hardware-centric focus at the time. The founding Parametricism
just a “put together” board game aggregating a set of principles he defined in his book are still today the Following such conclusive research, architects and
basic assembly rules and processes. The monotony of bedrock of software programming: concepts like the industry at large actively pushed these inventions In the world of parameters, both repetitive tasks and
the resulting designs rapidly trivialized the theory, and recursions, object-oriented programming as well as to the state of innovations. Frank Gehry certainly was complex shapes could possibly be tackled, when
the constructive weakness of its assembly systems their application to design have represented a radical the most vibrant advocate of the cause. For him, the rationalizable to simple sets of rules. The rules could
finally discouraged architects. Nevertheless, through move forward. Following this momentum, an entire application of computation could drastically relax the be encoded in the program, to automate the time-
 consuming process of manually implementing them. parameters to its users. As the software is released, resulting in a specific building shape. Hadid’s designs From Autodesk Revit -the major BIM software today-
This paradigm drove the advent of Parametricism. Geisberg perfectly summed up perfectly the parametric are the perfect examples to this day of the possible to Sutherland’s SketchPad, we see a single common
In few words, if a task can be explained as a set of ideal: quantification of architectural design, into arrays of thread: the explicit utilization of parameters as the
commands given to the computer, then the designer’s parameters. Her work, however, would have not been driving force of design.
task would be to communicate them to the software possible without Grasshopper, software developed by
“The goal is to create a system that would be flexible
while isolating the key parameters impacting the result. David Rutten in the year 2000’s. Designed as a visual However, the parametrization of design has proven
enough to encourage the
Once encoded, the architect would be able to vary the programming interface, Grasshopper allows architects over the past 10 years to have reached a plateau, both
engineer to easily consider a variety of designs. And the
to easily isolate the driving parameters of their design technically and conceptually. Parametric modeling
parameters and generate different possible scenarios: cost of making
different potential shapes, yielding multiple design while allowing them to tune them iteratively. The failed to account for (1) the compounded effect of
design changes ought to be as close
outputs at once. to zero as possible. “ multiple variables at once, (2) the imperative of
space organization and style over strict efficiency,
Now that the bridge between design and computation (3) the variability of scenarios, and finally (4) the
In the early 1960s, the advent of parametrized
was built thanks to Sutherland and Geisberg, a new computational cost of simulations. Independently
architecture was announced by Professor Luigi Moretti. from its technical shortcomings, parametric design is
His project “Stadium N”, although theoretical initially, is generation of “parameter-conscious” architects could
thrive. As architects were becoming more and more flawed by its theoretical premise: Architecture could
the first clear expression of Parametricism. By defining be the result of a fixed number of parameters, that the
capable of manipulating their design using the proxy
19 driving parameters — among which the spectators’ architect could simply encode, as an abstraction, away
of parameters, the discipline “slowly converged”
field of view and sun exposure of the tribunes -, Moretti from its context, its environment, and its history. In
to Parametricism, as explained by P. Schumacher.
14 derived the shape of the stadium directly from the In his book, “Parametricism, a New Global Style for fact, Parametricism, when applied ‘by the book’, proved 15
variation of these parameters. The resulting shape, Architecture & Urban Design” Schumacher explicitly to neglect the immense complexity of space planning:
although surprising and quite organic, offers the first demonstrated how Parametricism was the result of a countless parameters and profound cultural & societal
Figure 4: Grasshopper by David Rutten
example of this new parametric aesthetic: organic in growing awareness of the notion of parameters within factors actually participate in the urban equilibrium.
aspect, while strictly rational as a conception process. the architectural discipline. simplicity of its interface coupled with the intelligence This deep reality, combining adjacent disciplines in a
of the built-in features continues today to power most systemic way, can today finally be addressed, as our
Bringing such principle to the world of computation
From the invention of parameters to their translation buildings’ design across the world and has inspired an profession encounters Artificial Intelligence.
will be the contribution of Ivan Sutherland, three years
later. Sutherland is the creator of SketchPad, one of the into innovations throughout the industry, we see a entire generation of “parametric” designers. Finally,
first truly user-friendly CAD software. Embedded at the handful of key individuals, who have shaped the advent beyond the short-term benefits of Grasshopper for Artificial Intelligence: a Statisitical Approach to
heart of the software, the notion of “Atomic Constraint” of Parametricism. This parametrization of architecture building design, a more profound revolution, driven Architecure
is best exemplified at first by Zaha Hadid Architects’ by parametrization and started in the early 2000s,
is Sutherland’s translation of Moretti’s idea of parameter.
work. Mrs. Hadid, an Iraqi architect trained in the UK, is still underway today: BIM (Building Information Artificial Intelligence is fundamentally a statistical
In a typical SketchPad drawing, each geometry was in
with a math background would found her practice, Modeling). Spearheaded by Philip Bernstein, then Vice approach to architecture. The premise of AI, that
fact translated on the machine side into a set of atomic with the intent to marry math and architecture President of Autodesk, the birth and refinement of BIM blends statistical principles with computation is a new
constraints (parameters). This very notion is the first through the medium of parametric design. Her designs has brought rationality and feasibility to a brand-new approach that can improve over the drawbacks of
formulation of parametric design in computer’s terms. would typically be the result of rules, encoded in the level within the construction industry. The underlying parametric architecture.
Samuel Geisberg, founder of the Parametric Technology program, allowing for unprecedented levels of control idea of the BIM is that every element in a building
Corporation (PTC), would later, in 1988, roll out Pro/ over the buildings’ geometry. Each architectural move 3D model is a function of parameters (“properties”) “Learning”, as understood by machines, corresponds
ENGINEER, first software giving full access to geometric would be translated into a given tuning of parameters, that drives each object’s shape and document them. to the ability of a computer, when faced with a
 complicated issue, first to grasp the complexity of the optimize adjacencies and light condition as the the idea of encoding a behavior, that the GENERATOR potential: networks and machine learning. Through
options shown to him and second to build an “intuition” user draws onto a modular grid. In fact, URBAN V would follow. However, bellow Negroponte’s work, or the utilization of a layered pipeline, also called network,
to solve the problem at stake. In fact, when coining distinguished two layers of information: implicit and Price’s prototypes, lied an unresolved issue: the actual a machine is now able to grasp higher complexities
down the concept of AI, John McCarthy, back in 1956, explicit. The implicit dimension is the one handled and intelligence of the algorithm. Although the interface then previously developed models. Such models can
has defined it as “using the human brain as a model deduced by the machine, while the explicit one is the and protocols were in place, the actual procedural be “trained”, or in other words, tuned for specific tasks.
for machine logic”. Instead of designing a deterministic dimension set by the user. This duality of information complexity of the core algorithms was still quite weak, More interesting even, is the idea embarked in one
model, built for a set number of variables and rules, in URBAN V is the direct translation of the machine- based on simple heuristics relationships. specific type of such models: Generative Adversarial
AI lets the computer create intermediary parameters, human complementarity wished by Negroponte. And Neural Networks (GANs). Theorized at first by Ian
from information either collected from the data or it is within the set of implicit parameters, that the The design of intelligent algorithms, also called AI, Goodfellow, researcher at Google Brain, in 2014, this
transmitted by the user. Once the “learning phase” “intelligence” — in other words, the AI- built within actually found a renewed interest at the beginning model offers to use networks to generate images, while
achieved, the machine can generate solutions, not the machine would find its expression. Corrections of the ’80s. The sudden increase in computational ensuring accuracy through a self-correcting feedback
simply answering a set of predefined parameters, but proposed by the computer, by tuning the implicit power and the steep increase of funding’s brought loop. Goodfellow’s research turns upside down the
creating results emulating the statistical distribution parameters, would be surfaced to the users as back the question of intelligence at the center of definition of AI, from an analytical tool to a generative
of the information shown to him during the learning AI’s investigation. Key to this period are two main agent. By the same token, he brings AI one step
phase. This concept is at the core of the paradigm revolutions: expert systems and inference engines. closer to architectural concerns: drawing and image
shift brought by AI. The partial independence of the The former corresponds to machines able to reason production. All in all, from simple networks to GANs,
machine to build its own understanding of the problem, based on a set of rules, using conditional statements. a new generation of tools coupled with increasingly
coupled with its ability to digest the complexity of a An actual breakthrough at the time. The later, best cheaper and accessible computational power is today
16 set of examples, turns upside down the premise of exemplified by the Cyc Project, developed by Douglas positioning AI as an affordable and powerful medium. 17
Parametricism. Since not all rules & parameters are Lenat, were involving machines geared towards If Negroponte’s or Price’s work were almost empty
declared upfront explicitly by the user, the machine inference reasoning. Using a knowledge base (a set of true machine intelligence, nowadays architectural
can unexpectedly unveil underlying phenomena and of truth statements), an inference machine would be software can finally leverage such possibility.
even try to emulate them. It is a quantum leap from able to deduce the truthfulness of a new statement
the world of heuristics (rule-based decision making) as compared to its knowledge base. It is not until the Although the potential AI represents for Architecture
to the world of statistics (stochastic-based decision early 90’s, and the mathematization of AI that the is quite promising, it still remains contingent on
making). field would bring truly promising results. The advent designers’ ability to communicate their intent to the
Figure 5: GENERATOR, Cedric Price of a new type of models would definitely reveal AI’s machine. And as the machine has to be trained to
become a reliable “assistant”, architects are faced
The penetration of Artificial Intelligence in the suggestions. To an ill-placed set of rooms, URBAN V with two main challenges: (1) they have to pick up an
architectural field was already forecasted early on would notify the user: “TED, MANY CONFLICTS ARE adequate taxonomy i.e. the right set of adjectives that
by a few theorists, who, before us, saw AI’s potential OCCURRING”. A few years later, Cedric Price, then can translate into quantifiable metrics for the machine
for architectural design. Far from crafting intelligent Professor at the chair of Architecture at Cambridge and (2) they must select, in the vast field of AI, the
algorithms, these precursors designed and speculated University, invented the GENERATOR (1976). proper tools and train them. Those two preconditions
on the potential of such systems. As URBAN II was Acknowledging Negroponte’s work, Price used the will eventually determine the success or the failure of
released by Negroponte and his group, the idea of AMG’s work on AI and pushed it further, investigating AI-enabled architecture.
a “machine assistant” was already well underway. the idea of autonomous ever-changing building, that
URBAN V, a later version, would assist the designer, would “intelligently” respond and adapt to users’
Figure 6: GAN Typical Architecture
by adapting rooms layout– defined as blocks — to behaviors. For Price, under the term “intelligent” lies
 Modularity Computational Design Parametricism Artificial Intelligence
Christopher Alexander
Notes on the Synthesis of Form (’64) J. Weizenbaum at MIT AI Lab
Gropius A Pattern Language (’68) ELIZA, First Chatterbot (’66) L. Moretti
Stadium N (’60)
“Baukasten” Concept N. Negroponte
“Computers for Architectural Design?”
Dartmouth Conference “The Architecture C.Price I. Sutherland & SketchPad
Le Corbusier Buckminster Fueller Artificial Intelligence Machine” (’70) GENERATOR (’76) Concept of “Atomic Constraints” (’63) I. Goodfellow
Generative Adversarial
Le Modulor (’45) Dymaxion House (’46) Invention (’56) Architecture D. Lenat P. Schumacher Neural Networks (’14)
Machine Group Cyc Project (’84) Parametricism - A New Global Style
Urban II and V for Architecture and Urban Design
Inventions & Theories
18 19
1930

1960

1990
Innovations

2010
Robert W. McLaughlin Moshe Safdie P. Hanratty & PRONTO Autodesk AutoCAD F. Gehry, J. Glymph & CATIA Zaha Hadid P. Bernstein & REVIT D. Rutten & GRASSHOPPER
Winslow Ames House (’33) Habitat 67 (’67) First CAD/CAM Software (’59) First Architectural CAD (’77) Guggenheim Bilbao (’93-’97) Vitra Fire Station (93) First Release : 2000 First Release : 2007 Architecture + AI
 “Relevance” compounds the “constructive
Variety x Relevance B. A Continuous Progress feasibility”, i.e. the workability of the designs
and their “architectural quality” including
Variety Relevance optimal program organization, space layout
and contextual fit.
Modularity, Computational Design, Parametricism,
and Artificial Intelligence: this four-period sequence Ultimately, the combination of Variety (Quantity X
reflects the chronology of the progress which, Singularity) and Relevance (Constructive Feasibility X

Achitectural Quality
step-by-step, has been shaping and refining the Architectural Quality) creates a framework which (1)
x

Variety
Quantity

20 = architectural means & methods. We want to see in maps out and contrasts the respective positioning of 21
such momentum a form of “continuous progress” as our four periods -Modularity, Computational Design,
the one experienced in the industry at large, rather Parametricism, and Artificial Intelligence and (2)
than a series of unrelated disruptions. From there, an clearly evidences the culminating point of progress
appropriate set of matrices has helped us mapping AI represents for our discipline.
Singularity Constructive Feasibility this dynamic.
Relevance
First, to evidence our claim, we posit here that
Architecture can be understood as a process of
generating designs one can describe through two
dimensions: on one side the diversity of the output
produced or “Variety” and on the other side the This display, although directional and qualitative, is a
applicability of the designs or “Relevance”. powerful grid to represent the concept which lies at
the heart of our thesis. In summary, the dynamic of
“Variety” is contingent upon two underlying this continuous progress has been triggered by the
metrics: the “quantity of designs” sizing the limit of each movement at a certain point in time,
volume of options created and the “singularity exacerbated by the competition of the new one
of designs” measuring their respective disparity. coming in.
 Variety Variety x Relevance Since its inception, modular conception has proved to be a
C. Closing Remarks
highly constraining system, yielding low variety of potential
design options. Although such options were easily meeting We are today faced with a fantastic challenge:
construction feasibility criteria, their architectural quality bringing AI to the world of architectural design. We
A A got questioned early on.

A
know we will have to resist to preconceived ideas and
natural fears. Let’s make it clear: we do not believe
P P

P
that AI will ever automate the architect’s intuition
and will substitute to his/her sensitivity. We consider

Variety
C C

C
that, in the foreseen future, humans will continue
With computation design, designer could finally afford to using the machine as their tool; not the other way
escape the rigidity of the grid, to design feasible buildings, around. However, we are also convinced that the
with actual singularity. Conception software, allowing to inception of an AI-powered “intelligent assistant” is

y
rit
Qu

M M

la
solve complex shapes, would help drive down conception
a

gu
M
nt

a game changer within our reach. And, as previously

ity

Sin
costs and generate more design iterations.
documented, we consider it conceptually fits in a
technological continuum, away from the simplistic
22 Relevance extension of the tabula rasa theory. It is why we are 23
quite confident that the architectural community
while recognizing the radical breakthrough AI
Parametricism brought even more control over organic
represents, will carefully study, test and experience
shapes, increasing constructive feasibility. By systematiz-
ing their geometry, entire buildings could be discretized
its applications further.

lity
ua
into buildable elements, and resolved assembly systems.

lQ
Meanwhile, it is the attempt of our thesis, expressing

ra
The parametric style, however, got rapidly trivialized, gen-

tu
ec
our belief that AI can balance design efficiency and
hit
erating generic and repetitive patterns spaces leading to
Ac poor architectural quality. organicity in an unparalleled way. We want to simply
C A demonstrate here that AI-powered space design
can help optimally combine hard sciences, such

A
C
P as engineering, math & data science, and softer-
Co

sciences, including design, architecture & planning.

ns

P
tru
ct

M We offer in the next chapters a framework with the

ive

Artificial intelligence finally carries the promise of embark-

Fe

intent to shape this challenging topic and bring early

a

M
sib

ing all the benefits of previous generations, while escaping

ilit

evidence of its materiality. Our hope is to jump-start

y

the legibility and generic style of Parametricism. It is the

ultimate push towards greater architectural quality and, to the discussion and lay down the premise of machine-
Relevance us the springboard of a brand-new era. assisted architectural design.
 architectural design. We postulate that its utilization
can enhance the practice of the architectural
D. Framework discipline. This field is as recent as it is experimental
and yields to this day surprising results. Our hope is to
be able to train it to draw actual building floor plans.
Our work finds itself at the intersection of
Architecture and Artificial Intelligence. The former Then, we will come up with a robust analytical
is the topic, the latter the method. Both have been framework to qualify and classify the generated
simplified into clear & actionable categories. floor plans. Ultimately, our goal is to organize the
results of our GANs, to offer the possibility for the
Architecture is here understood as the intersection user to browse seamlessly through the variety of
between Style and Organization. On one hand, created design options. To that end, the quantity and
we consider buildings as vectors of a cultural ubiquity of tools offered by Data Science will prove to
significance, that express through their geometry, be valuable to our investigation.
24 taxonomy, typology, and decoration a certain style. 25
Baroque, Roman, Gothic, Modern, Contemporary: as Through this dual lens, at the crossroad of Style &
many architectural styles that can be found through Organization, Qualification & Generation, we lay
a careful study of floor plans. On the other hand, down a framework that organizes the encounter of
buildings are the product of engineering and science, Architecture & AI.

Organization
answering to strict frameworks and rules -building
codes, ergonomics, energetic efficiency, egress,

Style
program, etc — that can be found as we read a floor
plan. This organizational imperative will complete our
definition of Architecture and drive our investigation.
Qualify

Artificial Intelligence will be employed, using two

of its main fields of investigation — Analytics and
Generative Adversarial Networks — as an investigative Generate
tool.

At first, we will dive into the topic of Generation. Using

GANs, we offer to educate our own AI systems to Framework Matrix | Source: Author
 II
AI & Generative Adversarial Neural Networks

Generate Generative Adversarial Neural Networks 

GANs- are here our weapon of choice. Within the
— or

field of AI, Neural Networks stands as a key field of

investigation. The creative ability of such models
has been recently evidenced, through the advent
of Generative Adversarial Neural Networks. As any
The design of architectural floor plans is at the core machine-learning model, GANs learn statistically
of the architecture practice. Its mastery stands as significant phenomena among data presented
26 the gold standard of the discipline. It is an exercise to them. Their structure, however, represents 27
that practitioners have overtime relentlessly tried a breakthrough: made of two key models, the
to improve through technology. In this first part, Generator and the Discriminator, GANs leverage
we dive into the potential of AI applied to floor plan a feedback loop between both models to refine
generation, as a mean to push the envelope even their ability to generate relevant images. The
further. Discriminator is trained to recognize images from
a set of data. Properly trained, this model is able
Using our framework, to tackle floor plans’ style to distinguish between a real example, taken out
and organization, we lay down in the following of the dataset, from a “fake” image, foreign to the
chapter the potential of AI-enabled space planning. dataset. The Generator, however, is trained to create
Our objective is to offer a set of reliable and robust images resembling images from the same dataset.
Typical GAN Architecture tools to both evidence the potential of our such an As the Generator creates images, the Discriminator
approach and to test our assumptions. provides him with some feedback about the quality
of its output. In response, the Generator adapts, to
The challenge is here threefold: (1) choosing the produce even more realistic images. Through this
right toolset, (2) isolating the right phenomena to feedback loop, a GAN slowly builds up its ability
be shown to the machine and (3) ensuring that the to create relevant synthetic images, factoring in
machine “learns” properly. phenomena found among observed data.
 Training Start

Representation & Learning

If GANs represent a tremendous opportunity for

us, knowing what to show them is crucial. We have
here the opportunity to let the model learn directly
from floor plan images. By formatting images, we
can control the type of information that the model
will learn. As an example, just showing our model the
shape of a parcel and associated building footprint will
yield a model able to create typical building footprints
given a parcel’s shape. To ensure the quality of the
outputs, we will use our own architectural “sense” to
28 curate the content of our training sets: a model will 29
only be as good as the data we give him, as architects.

Here on the left, we illustrate a typical training

sequence: this sequence, realized over the course
of a day and half of training, displays how one of
our GAN-models progressively learns how to layout
rooms and fenestration for housing units.

Although the initial attempts are imprecise and

confuse, after 250 iterations the machine builds for
itself some form of intuition.

Livingroom Bedroom Bathroom

Corridor Kitchen Closet

Training End
Training Sequence, Model II | Source: Author
 rooms across a single-family home footprint. Peters’
work turns an empty footprint, into programmatic
A. Precedents patches of color, without specified fenestration.

Regarding GANs as design assistants, Nono Martinez’

Since the early work of Goodfellow et al. in 2014 [1], thesis [6] at the Harvard GSD in 2017 investigates the
the field of Generative Adversarial Nets (GAN) has idea of a loop between the machine and the designer
expanded, from simple hand-drawn digits generation to refine the very notion of “design process”. Martinez
to more complex taxonomies generation. Isola et al. in trains models for specific sketching tasks and offers
November 2018 set a key precedent in this field [2], by an interface through which users can solicit each
enabling image-to-image translation. In their model, model at different moments of the drawing process.
Pix2Pix, pairs of images are fed to a GAN architecture,
while the network learns the proper mapping from Our work expands on these precedents and offers to
one image to the other. Part of their work investigates nest 3 models (footprint, program, and furnishing)
building façade generation and opens the door to to create a full “generation stack” while improving
architectural design using GAN.Professor Andrew Witt results quality at each step. By automating multi-
30 expands on this work in his exhibition QUILTING units processing, our work then scales to entire 31
By enlarging the final layers of Pix2Pix, Witt creates buildings generation, and masterplan layouts. We
larger facade designs, showcased as one linear endless further offer an array of models dealing with style
animation of an urban skyline. transfer. Finally, our contribution adds a rigorous
[5]
framework to parse and classify resulting outputs,
[6]
Remaining within the realm of Architecture, floorplans enabling users to “browse” consistently through
design using GAN is first studied by Zheng and Huang generated options.
in 2018 [3]. Using Pix2PixHD [4], the authors propose
[1] Ian J. Goodfellow, Jean [4] Ting-Chun Wang, Ming-Yu Liu,
to use GANs for floorplan recognition and generation. Pouget-Abadie, Mehdi Mirza, Bing Jun-Yan Zhu, Andrew Tao, Jan Kautz,
Xu, David Warde-Farley, Sherjil Ozair, Bryan Catanzaro. 2018. High-Resolu-
Floorplan images processed by their GAN architecture Aaron Courville, Yoshua Bengio. 2014. tion Image Synthesis and Semantic
are translated into programmatic patches of colors. “Generative Adversarial Networks”. Manipulation with Conditional GANs.
arXiv:1406.2661. CVPR. arXiv: 1711.11585v2.

[2] Phillip Isola, Jun-Yan Zhu, [5] Nathan Peters. 2017. Master Thesis:
Inversely, patches of colors in their work are turned into Tinghui Zhou, Alexei A. Efros. 2017. “Enabling Alternative Architectures:
drawn floorplans. If the position of openings and rooms ”Image-to-Image Translation with Collaborative Frameworks for
Conditional Adversarial Networks”. Participatory Design”.Harvard Graduate
is specified by the user, the elements actually being laid CVPR. arXiv:1611.07004. School of Design, Cambridge, MA.
out by the network are the furniture. The same year, [3] Hao Zheng, Weixin Huang. 2018. [6] Nono Martinez. 2016. “Suggestive
[3] [3] Nathan Peters [5] in his thesis at the Harvard Graduate “Architectural Drawings Recognition Drawing Among Human and Artificial
and Generation through Machine Intelligences”, Harvard Graduate
School of Design tackles the possibility of laying out Learning”. Cambridge, MA, ACADIA. School of Design, Cambridge, MA.
 B. Organization

In this section, we offer a multi-step pipeline,

integrating all the necessary steps to draw a floorplan.
Jumping across scales, it emulates the process taken 33
32
by an architect and tries to encapsulate each step
into one specific model, trained to perform each
given operation. From (I) the parcel to the building
footprint, (II) from the footprint to a room split, (III)
from a room split to a furnished one, each step has
been carefully engineered, trained and tested.

Generation Pipeline, Model I to III | Source: Author At the same time, by dividing the pipeline into
discrete steps, the system allows for the architect’s
intervention between each model. As each model
generates multiple options at each step, the
architect’s ability to select the output of a model and
edit it before transferring it to the next model keeps
him/her in control of the design process. Its input
shapes the decisions made by the model, therefore
achieving the human-machine interaction expected.
I II III
34 35
 I Footprint
36 37
The first step in our pipeline tackles the challenge of
creating an appropriate building footprint for a given
parcel geometry.

To train this model, we used an extensive database

of Boston’s building footprints and were able to
create an array of models, each tailored for a specific
property type: commercial, residential (house),
residential (condo), industrial, etc.
 Each model is able for a given parcel, to create a set
of relevant footprints, resembling in dimension and
38 style the type it was trained for. 39
9 examples, using the residential model are shown
here on the left.

Results: Generated Footprints | Source: Author

 II Room Split

The layout of rooms across a building footprint is

the natural next step. Being able to split a given
40 floor plan, while respecting meaningful adjacencies, 41
typical room dimensions and proper fenestrations
is a challenging process, that GANs can tackle with
surprising results.

Using a dataset of around 700+ annotated floor

plans, we were able to train a broad array of models.
 Each model is geared towards a specific room count
and yields surprisingly relevant results once used on
empty building footprints.

We display here on the left some typical results.

42 43

More results can be found at the following address:

http://stanislaschaillou.com/thesis/GAN/unit_opening_results/

Results: Generated Program & Fenestration | Source: Author

 III Furnishing

This last step brings the principle of generation to its

most granular level: the addition of furniture across
44 space. To that end, we trained at first a model to 45
furnish the entire apartment all at once.

The network was able to learn, based on each room

program, the relative disposition of furniture across
space, and the dimensions of each element.

For the sake of time and accuracy, we simplified our

models to restricted types of furniture.
 The geometry of furniture is not always perfectly
accurate. However furniture types are legible, and
their relative position in space is reasonable.

We display here on the left some typical results.

46 47

More results can be found at the following address:

http://stanislaschaillou.com/thesis/GAN/unit_furnishing_results/

Results: Furnished Units | Source: Author

 Room Rendering
If the previous results can give a rough idea of
potential furniture layouts, the quality of the resulting
drawings is still too fuzzy. To further refine the
48 output quality, we have trained an array of additional 49
models, for each room type (living room, bedroom,
kitchen, etc…).

Each model is only in charge of translating a color

patch added onto the plan, into a properly drawn
piece furniture. Furniture types are encoded using a
color code. We display here on the left the results of
each model.

Rendered Rooms: Bathrom, Kitchen, Livingroom, Bedroom | Source: Author

Room Rendering Interface

Simple User Interface for Bathroom, Livingroom,
Bedroom & Kitchen Floorplan Rendering

http://stanislaschaillou.com/thesis/GAN/canvas
 Going Further

If generating standard apartments can be achieved

using our technic, pushing the boundaries of our
models is the natural next step. GANs can, in fact,
offer quite remarkable flexibility to solve seemingly
highly constrained problems. In the case of floor
50 Generation Pipeline, Multiple Units | Source: Author plans layout, as the footprint changes in dimension 51
and shape, partitioning and furnishing the space by
hand can be a challenging process. Our models prove
here to be quite “smart” in their ability to adapt to
changing constraints.

Our ability to control the units’ entrance door and

windows position, coupled with the flexibility of our
models allows us to tackle space planning at a larger
scale, beyond the logic of a single unit. By adding
some simple algorithms to our generation pipeline
Model II Results | Source: Author Model III Results | Source: Author
(here on the right), we enable users to scale their
work to the design of entire buildings, containing
multiple apartment units.

In the next pages, we display the results of our

improved generation methodology, as it scales to
entire buildings, and neighborhoods.
52 53
54 55
56 57
58 59
 Then, we technically believe that the sequential
nature of the application will facilitate its
Organization manageability and foster its development. The
ability to intervene throughout the generating
Takeaways process is a fundamental dimension: as each step
of the pipeline represents a distinct portion of
architectural expertise, each model can be trained
60 Beyond the strict development of a generation independently, opening the way to significant 61
pipeline, we lay down in this chapter certain intuitions improvements and experimentation in the near
and concepts that will shape the rest of this thesis. future. Indeed, improving this entire pipeline end-
to-end could be a long and cumbersome task, while
First, we are directionally convinced that our amending it step by step remains a manageable
ability to design the right pipeline will condition process, within the reach of most architects and
AI’s success as a new architectural toolset. Our engineers in the industry.
choice for the “Grey Boxing” approach, as introduced
by Prof. Andrew Witt in Log, will likely secure the Finally, we hope our framework will help address
best potential results. This method contrasts with the endless breadth and complexity of the models
the “black box” approach, that only allows users to to be trained and those used in any generation
input information upfront, and to get finished design pipeline. Tackling parcels-footprint-room split-etc.,
options at the end of the process, without any control as we do is one possible approach among, we believe,
over the successive generation steps. To the contrary, a large set of options. To encapsulate the necessary
by breaking out our pipeline into discrete steps, “Grey steps of space planning, the key is more the principle
Boxing” permits the user to intervene all along the than the method. And with the growing availability of
way. His tight control over the machine is his ultimate architectural data, we encourage further work and
guarantee of the design process quality. open-minded experimentation.
 C. Style

As we refine our ability to generate floorplans, we

raise the question of the bias intrinsic to our models
Baroque Manhattan Unit
62 and offer here to extend our study beyond the 63
simple imperative of organization. We investigate
architectural style learning, by training and tuning
an array of models on specific styles: Baroque, Row
House, Victorian Suburban House, & Manhattan Unit.

Beyond the simple gimmick of each style, our

study reveals the deeper meaning of stylistic: more
than its mere cultural significance, style carries a
fundamental set of functional rules that defines a
clear mechanic of space and controls the internal
Row-House Suburban Victorian organization of the plan. In his part, we will try to
evidence the profound impact of architectural style
on the composition of floorplans.
 Bias & the Emergence of Style

Taking into consideration a batch of generated

units, using our initial pipeline, we start to notice
some amount of intrinsic bias to our model: the
64 internal wall structure is consistently laid out as an 65
orthogonal system of partitions, disregarding the
potential orientation of the units’ facades (see image
here below). At the same time, the program layout
is also consistently set up such that “serving” spaces
-bathroom, toilets, closets, kitchen- are packed at
the back of the plot, while the odd geometry of the
facade gets absorbed by over-dimensioned living
rooms and bedrooms.

These characteristics can, in fact, be found all across

Bias & Style Among Generated Results | Source: Author
our initial training set. We understand here this
reality as the literal translation of a concept central
to the architectural discipline: style.
 Four Styles

Instead of preventing this bias, striving to create a

generic or objective plan generator, we will rather
embrace it and study its presence to eventually use
it to our advantage.

To that end, we choose to broaden our investigation

and extend it to architectural style learning. We
Baroque
create a new pipeline, enabling the conversion of
floorplans, from one style to another; here on the
left, from Modern to Baroque.

This example reveals all the more the deeper

66 meaning of architectural style: we notice that the 67
translation through model A & B is not simply a new
make-over of the existing figure wall, but rather
a profound remodeling of internal structures and
spatial organizations.
Manhattan Unit Suburban Victorian

In fact, we evidence here what Farshid Moussavi

coined down as the “Function of Style” in her book.
Each style, beyond its cultural significance, handles
space differently and reacts specifically to similar
constraints.
Style Transfer Pipeline | Source: Author
To investigate architectural style learning, we have
trained and tuned an array of models on specific
Row-House styles — Baroque, Row House, Victorian Suburban
House, & Manhattan Unit — able to emulate each
particular architectural style. The results are
displayed in the following pages.
68
Baroque Style Generated 69
Training Set Units
70
Manhattan Style Generated 71
Training Set Units

Row-House Style Training Set Generated Units

72
Victorian Style Generated 73
Training Set Units
74
Row-House Style Generated 75
Training Set Units
76
Baroque Style Manhattan Style 77
Physical Model Physical Model
78
Victorian Style Row-House Style 79
Physical Model Physical Model
 Among the generated units we can identify some
clear patterns within each style. This “behavior”
proper to each model is a direct translation of
each style’s mechanic. More than applying a simple
texture across each apartment unit, each model has
captured a set of characteristics and rules.

To move beyond this simple observation, we offer

to map out each style’s abilities. To each model
corresponds a set of strengths & weaknesses, and
coining them down will allow us to truly assess the
80 functional reality of each style. In addition, our hope is 81
to expand our understanding of our models’ abilities
to allow us later to use each one purposely, given a
new set of constraints, and functional requirements.

In clear we propose a six-axis graph, reflecting a

given model’s ability to handle six specific types of
condition: Depth, Compactness, Single-Orientation
or Multi-Orientation (number of facades), Acute
Angle (sharp geometry of the boundary), Program
Spectrum (breadth of the program).

After thoroughly testing our four models, we propose

the graphs here on the left.
Characteristic Graphs | Source: Author
 Application

Finally, we have brought all these intuitions together

in a final architectural project: a large-scale housing
82 development located in Manhattan’s Lower East Side. 83
The complex geometry of the parcel forces a certain
amount of complexity on our design. As a result of
our massing (on the far left), we obtain a catalog of
380 one-of-kind apartment units (here on the left).

Catalog of All Units | Source: Author

 We first attempt to process an entire floor, each time
using a different style. The results displayed here
n the left reveal once again the necessity of using
styles cautiously, respectful of the constraints and
84 85
the specificity of each context. If certain units are
successfully laid out using a certain style, others fail
to find a proper internal organization. By alternating
styles across our catalog of apartment units, we hope
to find an appropriate answer to each specific spatial
condition.

15th Floor Processed Under Each Style: Baroque (Far-Left), Manhattan (Center-Left),
Row-House (Center-Right), Victorian (Far-Right) | Source: Author
 Knowing each GAN-model’s strengths and
weaknesses, each style’s potential and short-
comings, we now process every apartment using
the most suited model. Each floor is turned into
a patchwork of styles. Our goal becomes then to
compose our “mosaic” picking for each tile -each
unit- the most reasonable model that will best 87
86
handle the constraints. Out of this selection process,
we have isolated some resulting options, displayed
here below.

We then narrow down our exploration, precise the

selection of units and styles across all floorplates, to
finally assemble our final design. Here on the left are
three typical floors.

15th Floor 16th Floor 4th Floor

 models, one for laying out load-bearing elements,
and one for adding partitioning walls.

The Strict Imperative of Efficiency

Both the exuberance of each style and the level

of freedom given to our models do not address a
Assumptions & Limitations concern common in our discipline: space efficiency.
However, our primary concern here is to maximize
the expressiveness of each style, to let each model
unfold its mechanic to showcase its “personality”.
We have in fact turned styles into functional tools,
able to address specific conditions all across our Note: To reconcile GANs with efficiency, we posit that
development. their outputs constitute a tremendous initialization
for standard optimization technics. Parametricism’s
88 However, as the above plans suggest, we set aside typical pitfall was to setup a too-broad problem space, 89
certain constraints, and make clear assumptions. We coupled with a random initialization. Optimizations
would like to clarify these. run on these settings would often end up converging
on locally minimal solutions. The intuition behind our
The Structure GANs brings a whole new quality of initialization, that
narrows down significantly the problem space while
The structure is left to the structural cores, and the setting up a good-enough initial solution.
tensions cables running along the facade. The plan
is therefore uninterrupted by vertical loads and The Massing
allows our algorithms to generate freely each unit’s
partitioning system. The notion of Massing refers to the outer-shape of
our building. The irrational form of the above design
Note: Looking back to our generation pipeline, a is here meant as a trigger for complexity, creating an
potential improvement would take the position of actual challenge for our models. To a more rationale
load-bearing walls & columns as inputs for Model I. massing would correspond more tamed & realistic
In such a way, our pipeline would allow designers to apartment units designs. Our focus here remains
control the building structural system. Alternatively, to showcase an extreme case and test our models’
Model I could be unpacked into two successive limits.
 Catalog

We turn finally to the catalog of units generated,

found across our building (here on the left). The
coherence and richness of the resulting designs
are striking. Moreover, the “intelligence” or formal
flexibility displayed in the generated apartments
90 91
further evidences the validity of the approach: GAN-
models can indeed encapsulate some amount of
architectural expertise & stylistic that can be later
used, depending on the set of constraints at play.
The “personality” of each model described in Part II
is clearly legible among each subset.

To conclude this section, we offer in the following

pages a sequence of key shots, taken across our
catalog. To the strict descriptive nature of the plans
(left), we associate each time an image of the interior
atmosphere (right), as a way to reconcile our process
with the more experiential nature of Architecture.
92
93

Manhattan
94
95

Manhattan
96
97

Manhattan
98
99

Manhattan
100
101

Manhattan
102
103

Baroque
104
105

Baroque
106
107

Baroque
108
109

Baroque
110
111

Baroque
112
113

Row-House
114
115

Row-House
116
117

Row-House
118
119

Row-House
120
121

Row-House
122
123

Row-House
124
125

Victorian
126
127

Victorian
128
129

Victorian
130
131

Victorian
 of these implicit rules. Being able to encapsulate
each style could allow us to go beyond the study of
precedents, and complement it by unpacking the

Style
behavior of GAN-models such as the ones trained
here. Their ability to emulate some of the unspoken
rules of architecture could allow us to address the
Takeaways “quality with no name” embedded in buildings that
132 Christopher Alexander defines in his book The 133
Timeless Way of Building. AI is simply a new way to
If we can think of floorplans first as compositions, study it.
before being strictly the product of engineering,
then studying the driving forces of the composition Finally, the inherent presence of style within each
is maybe where AI can offer us some meaningful GAN-model constitutes a key takeaway: far from
answers. Following this intuition, we have evidenced the promise of an agnostic & objective practice of
in this chapter that architectural styles carry at a generative design, it seems that style permeates
deeper level an implicit mechanic of space, that irrevocably the very essence of any generative
significantly impacts any floorplan’s composition. In process. In clear: Style is not an ancillary, superficial
clear, there are spatial consequences to choosing a or decorative addendum. Style is at the core of
given style over another. the composition. Recognizing this evidence is a
prerequisite to understanding what AI can bring
At a more fundamental level, we can think of styles as to Architecture. In other words, there will be no
being the by-product of architectural history. If there agnostic-AI for Architecture, no style-less machine,
is within each style a deeper set of functional rules, no objective generative design. On the contrary,
then studying architectural history could potentially each model or algorithm will come with its flavor, its
be about understanding the evolution over time personality, its know-how.
 III
Qualify

“Mal nommer les choses c’est ajouter au

malheur du monde” – Albert Camus

“Failing to name things adds to the World’s disarray”

Access Code Source

Footprint Program The study of Architecture requires us to agree
134 on terms and adjectives to make sure computers 135
metrics can eventually grasp the different facets
of building design. This imperative becomes most
crucial, as the Architect plans to rely on computers
to help their design process. Transforming adjectives
into quantifiable metrics and encoding them
becomes the necessary bridges between the human
Connectivity Circulation
and the machine.

To that end, we offer in this chapter six key metrics,

qualifying six essential aspects of floorplan design:
Footprint, Program, Orientation, Thickness & Texture,
Connectivity, and Circulation. These metrics work
together as a comprehensive framework, addressing
Orientation Thickness & Texture both the stylistic and organizational dimensions of
floorplans. Each has been developed as an algorithm,
thoroughly tested, and released as an open source
set of tools.
 Footprint

The shape of a building is the simplest and most

intuitive proxy to qualify its style. The “Footprint”
metric analyzes the shape of a floorplan perimeter
and translates it into a histogram. This descriptor,

Access Code Source

while encoding the shape of a building, can translate
136 common adjectives — “thin”, “bulky”, “symmetrical”, 137
etc. — used by architects into numerical information,
in order to communicate with a computer about
building shapes.

From a technical standpoint, this metric uses polar

convexity to turn a given outline into a list of discrete
values (vectors) that can then be compared to other
floorplans. We use a polar array of lines, stemming

Area
from the center of the plan, to extract the area of the
plan captured by each slice of space obtained. This
methodology has proven to yield satisfactory results,
Angle as shown in the queries here on the left. This technic
can also be employed to qualify indoor spaces’ shape
as well as building perimeter’s geometry.
 Query Result

Query Result

138 139

Query Result Query Result

 Query Result Programs

Program

Access Code Source

The program of a building, displaying the type of
140 rooms it contains, is a major driver of its internal 141
organization. Capturing this reality is central to
Query Result Programs
our approach. To describe the “mix” of rooms we
represent through a color code the list of rooms
contained in any given floorplan. This colored band
becomes the proxy, to describe the program. It
acts as a template, aggregating both the quantity
and the programmatic quality of the rooms within
the floorplan. It is an intuitive visual description for
humans which can translate into a reliable encoding
Query Result Programs
technic for machines.

From a technical standpoint, using this colored band

enables us to compute the programmatic similarities
and dissimilarities between any given pair of
floorplans. To visualize results, each plan is reported
as both a colored floorplan and one-dimensional
color vector of its program.
142 143
144 145
146 147
 Orientation

The orientation of walls in a plan is a valuable source

Access Code Source

of information. It can describe both the enclosure of
148 149
a plan (how secluded spaces are due to the walls’
presence) and the style of plan. In fact, using this
metric, we can easily differentiate a modern house-
pavilion from a gothic cathedral, simply by extracting
the histogram of the walls’ orientation.

From a technical standpoint, orientation extracts the

walls of a given floorplan and sums their length along
each direction of space, from 0 the 360 degrees. The
resulting list of values is an assessment of the overall
orientation of the plan. It can be averaged to get a
single descriptor or used as a vector to compare
across plans.
150 151
152 153
154 155
 Thickness & Texture

Thickness & Texture qualify the “fat” of the plan: its

Access Code Source

Wall Count

Variation
wall thickness and the variation of this thickness.
156 157
The thickness of walls across a plan, as well as the
Wall Thickness (in pixels) Wall Texture (Thickness Variation) geometry of the wall surface — Texture — can vary
drastically from one style to other. A Beaux Arts
Hall would display columns and indented thick walls
when a villa from Mies van der Rohe would display
thin rectilinear walls, which our metric would grasp
easily.

Wall Count

Variation
From a technical standpoint, this metric isolates all
the walls of a given plan and outputs a histogram of
Wall Thickness (in pixels) Wall Texture (Thickness Variation) wall thicknesses. At the same time, the algorithm
computes the variation of the thickness, to better
describe the wall texture (ie. flat walls versus
mullions).
 Wall Count

Wall Count
Variation

Variation
Wall Thickness (in pixels) Wall Texture (Thickness Variation) Wall Thickness (in pixels) Wall Texture (Thickness Variation)

Wall Count

Wall Count
Variation

Variation
158 159

Wall Thickness (in pixels) Wall Texture (Thickness Variation) Wall Thickness (in pixels) Wall Texture (Thickness Variation)
Wall Count

Wall Count
Variation

Variation
Wall Thickness (in pixels) Wall Texture (Thickness Variation) Wall Thickness (in pixels) Wall Texture (Thickness Variation)
 Wall Count

Wall Count
Variation

Variation
Wall Thickness (in pixels) Wall Texture (Thickness Variation) Wall Thickness (in pixels) Wall Texture (Thickness Variation)

Wall Count

Wall Count
Variation

Variation
160 161

Wall Thickness (in pixels) Wall Texture (Thickness Variation) Wall Thickness (in pixels) Wall Texture (Thickness Variation)
Wall Count

Wall Count
Variation

Variation
Wall Thickness (in pixels) Wall Texture (Thickness Variation) Wall Thickness (in pixels) Wall Texture (Thickness Variation)
 Connectivity

Access Code Source

162 Connectivity tackles the question of room adjacency. 163
The proximity of rooms to one another is a key
dimension of a floorplan. Moreover, their connection
through doors and corridors defines the existence
of connections between them. Connectivity
investigates the quantity and quality of such
connections, by treating them as a standard graph.

From a technical standpoint, by using the fenestration

from the plan, we can deduce the graph of existing
relationships among rooms. The Connectivity
metric then builds an adjacency matrix, reporting
these connections. A graph representation is finally
generated. Using this graph, we can compare
floorplans, taking into account the similarity of
connections among rooms.
164 165
166 167
 Circulation

Access Code Source

168 169
The circulation in floorplan captures how people
move across it. By extracting a skeleton of the
circulation, or in other words, a wireframe of the
circulatory network, we can both quantify and qualify
people’s movement across a floorplan.

From a technical standpoint, Circulation extracts

the skeleton of circulations of a given floorplan and
sums its length along each direction of space, from
0 the 360 degrees. The resulting histogram is an
assessment of the circulatory network geometry and
can be used to be compared against other floorplans’
circulation.
170 171
172 173
 V
Mapping & Browsing

Looking back at our GAN-models, each one actually

outputs multiple options at each step of our
generation pipeline. The designer is then invited
to “pick” a preferred option, modify it if needed,
before actioning the next step. Browsing through
the generated options however can be frustrating,
and time-consuming. To that end, the set of metrics
defined in the “Qualify” chapter can demonstrate
174 their full potential here and complement our 175
generation pipeline. By using them as filters, the user
can narrow down the range of options and find in a
matter of seconds the relevant option for its design.
This duality of Generation-Filtering is where the
value of our work gets all the more evidenced: we
provide here a complete framework, leveraging AI
while staying within reach of a standard user.

Once filtered according to a given criterion (Footprint,

Program, Orientation, Thickness & Texture,
Connectivity or Circulation), we provide the user with
a tree-like representation of her/his choice. At the
center is a selected option, and around it, its nearest
neighbors classified according to a user-selected
Footprint Tree criterion. The user can then narrow down the search
Similarity Tree, computed according to Footprint Metric and find its ideal design option, or select another
option within the tree, to recompute the graph.
http://stanislaschaillou.com/thesis/tree/convexity
176 177

“Lorem ipsum dolor sit amet, consectetur adipiscing

elit, sed do eiusmod tempor incididunt ut labore et
dolore magna aliqua. Ut enim ad minim veniam, quis
nostrud exercitation ullamco laboris nisi ut aliquip ex
ea commodo consequat.

Footprint Tree
Similarity Tree, computed according to Footprint Metric

http://stanislaschaillou.com/thesis/tree/convexity
178 179

Program Tree
Similarity Tree, computed according to Program Metric

http://stanislaschaillou.com/thesis/tree/program
180 181

Connectivity Tree
Similarity Tree, computed according to Connectivity Metric

http://stanislaschaillou.com/thesis/tree/graph
 Then, nesting models one after the other will ultimately
allow (1) to encapsulate relevant pieces of expertise,
(2) while designers will be able to intervene between
each model, thus achieving the back-and-forth
between humans and machines

However, AI is also the opportunity for our discipline

Conclusion to embrace a new set of investigative methodologies.
If we applied here our GAN-models to architectural
style learning, many other phenomena found in
Architecture can be studied. AI simply brings new
182 means to architectural research, allowing us to unpack 183
AI will soon massively empower architects in their the complexities found in the built environment.
day to day practice. As such potential is about to be We see here the possibility for rich results, that will
demonstrated, our work participates to the proof of complement our practice and address some blind spots
concept while our framework offers a springboard for of our discipline.
discussion, inviting architects to start engaging with AI,
and data scientists to consider Architecture as a field of Finally, Artificial Intelligence is the opportunity
investigation. However, today, our manifesto could be for Architecture to reinstate simplicity and clarity
summarized in four major points. as a driving principle of our discipline. As we craft
methodologies, a new ethos based on investigative rigor
Conceptually first, our belief is that a statistical and scientific rationality can offer a valid intellectual
approach to design conception shapes AI’s potential anchor to architectural research. Far from the confusion
for Architecture. Its less-deterministic and more- that sometimes prevails in o ur field, we have here the
holistic character is undoubtedly a chance for our opportunity to build a new sound foundation for
field. Rather than using machines to optimize a set of the architectural discourse. It is a necessity that will
variables, relying on them to extract significant qualities ultimately bring back Architecture where it belongs: as
and mimicking them all along the design process is a an Art & a Science, a product of intellectual sensibility,
paradigm shift. and technical rigor.
 Bibliography
Suggestive Drawing Among Human and Artificial In- Raster-to-Vector: Revisiting Floorplan Transforma-
Theoretical References telligences, Nono Martinez, Harvard GSD Thesis, 2016 tion, Chen Liu,Jiajun Wu,Pushmeet Kohli,Yasutaka Fu-
rukawa, 2017, Washington University, Deep Mind, MIT
Digital Architecture Beyond Computers, Roberto
Enabling Alternative Architectures : Collaborative
Botazzi, Bloomsbury
Frameworks for Participatory Design, Nathan Peters, Relational Models for Visual Understanding of
Harvard GSD Thesis, 2017 Graphical Documents. Application to Architectural
Data-Driven Design & Construction, Randy Deutsch,
Drawings, Llus-Pere de las Heras, 2014, Universitat
Wiley
Autonoma de Barcelona
Shape matching and modeling using skeletal con-
Architectural Intelligence, How Designers and Ar-
text, Jun Xie,Pheng-Ann Heng,Mubarak Shah, 2007,
chitects Created the Digital Landscape, Molly Wright Technical References University of Central Florida,Chinese University of
Steenson, MIT Press
Hong Kong
DANIEL: A Deep Architecture for Automatic Anal-
Architectural Google, Beyond the Grid — Architec- ysis and Retrieval of Building Floor Plans, Divya
Statistical segmentation and structural recognition
ture & Information Technology pp. 226–229, Ludger Sharma,Nitin Gupta,Chiranjoy Chattopadhyay,Sameep
for floor plan interpretation, Lluís-Pere de las Heras,
Hovestadt, Birkhauser Mehta, 2017, IBM Research, IIT Jodhpur
Sheraz Ahmed, Marcus Liwicki, Ernest Valveny, Gem-
ma Sánchez, 2013, Computer Vision Center, Barcelona,
Algorithmic Complexity: Out of Nowhere, Complex- Automatic Room Detection and Room Labeling
Spain
ity, Design Strategy & World View pp. 75–86, Andrea from Architectural Floor Plans, Sheraz Ahmed, Mar-
Gleiniger & Georg Vrachliotis, Birkhauser cus Liwicki, Markus Weber, Andreas Dengel, 2012, Uni-
Unsupervised and Notation-Independent Wall Seg-
versity of Kaiserslautern
mentation in Floor Plans Using a Combination of
Code & Machine, Code, Between Operation & Narra-
184 Statistical and Structural Strategies, Lluıs-Pere de 185
tion pp. 41–53, Andrea Gleiniger & Georg Vrachliotis, Generative Adversarial Nets, Ian J. Goodfellow,
las Heras, Ernest Valveny, and Gemma Sanchez, 2014,
Birkhauser Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David
Computer Vision Center, Barcelona, Spain
Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua
Gropius’ Question or On Revealing And Concealing Bengio, June 2014 ,Universite de Montreal
Path Planning in Support of Smart Mobility Appli-
Code in Architecture And Art, Code, Between Opera-
cations using Generative Adversarial Networks, Mo-
tion & Narration pp. 75–89, Andrea Gleiniger & Georg
hammadi, Mehdi, Ala Al-Fuqaha, and Jun-Seok Oh.  ,
Vrachliotis, Birkhauser Pix2Pix, Image-to-Image Translation with Condi-
2018
tional Adversarial Networks, Phillip Isola, Jun-Yan
Soft Architecture Machines, Nicholas Negroponte, Zhu, Tinghui Zhou, Alexei A. Efros
Automatic Real-Time Generation of Floor Plans
MIT Press
Based on Squarified Treemaps Algorithm, Fernando
Automatic Interpretation of Floorplans Using Spa-
Marson and Soraia Raupp Musse, 2010, PUCRS
The Architecture Machine, Nicholas Negroponte, MIT tial Indexing, Hanan Samet, Aya Soffer, 1994, Univ-
Press eristy of Maryland
Procedural Modeling of Buildings, Pascal Muller,Peter
Wonka,Simon Haegler,Andreas Ulmer,Luc Van Gool,
A Pattern Language, 1968, Christopher Alexander Parsing Floor Plan Images, Samuel Dodge,Jiu Xu
2015, ETH Zurich, Arizona State University
,Bjorn Stenger, 2016, Arizona State University, Rakuten
Notes on the Synthesis of Form, 1964, Christopher Institute of Technology
Generative Design for Architectural Space  Plan-
Alexander
ning, Lorenzo Villaggi and Danil Nagy, 2017, Autodesk
Project Discover: An Application of Generative De-
Research
Cartogramic Metamorphologies; or Enter the Rowe- sign for Architectural Space Planning, Danil Nagy,
Bot, Andrew Witt, Log #36 Damon Lau, John Locke, Jim Stoddart, Lorenzo Vil-
laggi, Ray Wang, Dale Zhao and David Benjamin, 2016,
Grey Boxing, Andrew Witt, Log #43 The Living, Autodesk Studio
 Digital Resources
Thesis Website Footprint Metric | Notebook
Algorithm & Applications of Footprint Metric
Main Website of Thesis Project
https://github.com/StanislasChaillou/
http://stanislaschaillou.com/expliquer
OpenPlan/blob/master/Footprint.ipynb

Github Repository | Github Connectivity Metric | Notebook

Main Code Repository of Thesis Project Algorithm & Applications of Connectivity Metric

https://github.com/StanislasChaillou/Open-
https://github.com/StanislasChaillou/OpenPlan
Plan/blob/master/Connectivity.ipynb

Connectivity Tree | Visualization Orientation Metric | Notebook

Similarity Tree, computed according to Connectivity Metric Algorithm & Applications of Orientation Metric

https://github.com/StanislasChaillou/
http://stanislaschaillou.com/thesis/tree/graph
OpenPlan/blob/master/Orientation.ipynb
186 187
Program Tree | Visualization Program Metric | Notebook
Similarity Tree, computed according to Program Metric Algorithm & Applications of Program Metric

https://github.com/StanislasChaillou/
http://stanislaschaillou.com/thesis/tree/program
OpenPlan/blob/master/Program.ipynb

Footprint Tree | Visualization Thickness & Texture Metric | Notebook

Similarity Tree, computed according to Footprint Metric Algorithm & Applications of Thickness & Texture Metric

https://github.com/StanislasChaillou/OpenPlan/
http://stanislaschaillou.com/thesis/tree/convexity
blob/master/Thickness_Texture.ipynb

Mapping Similarity | Notebook Circulation Metric | Notebook

Maps and Visualizations of floor plans similarities Algorithm & Applications of Circulation Metric

https://github.com/StanislasChaillou/OpenPlan/ https://github.com/StanislasChaillou/
blob/master/Mapping_Similarity.ipynb OpenPlan/blob/master/Circulation.ipynb
 erutcetihcrA + IA
188

hcaorppA weN a sdrawoT

9102 | DSG dravraH | uolliahC salsinatS