Significant Others

Architectural Form-Finding in the age of Parametric Modeling and Machine Learning

Kyle Steinfeld for ARCH 204a, Fall 2018

At several moments
throughout the 20th and 21st centuries,
architects have found it useful to imagine that
the artifacts of design thinking
themselves have a voice.

Due to recent developments in design technology,
this useful metaphor may soon become
a literal reality.

Creative Help
Gordon and Roemmele, USC 2015
An application that helps writers by generating suggestions for the next sentence in a story as it being written, based on a model trained on a corpus of twenty million English-language stories.

This proposal concerns new possibilities in creative human machine collaboration. In that spirit, I thought I would start the presentation with this: an on-line creative writing tool that completes an author's sentences with computer-generated ones that are based on a corpus of millions of young adult fiction novels.

Frei Otto: Form Finding and Path Systems

Tell me if you've heard this one.
Begin with a metal ring the size of a dinner plate.
Divide the circumference of this ring into some number of evenly spaced points.
Working with dry wool thread, connect each of these points to every other, but do not pull them taut. 
Instead, leave 10% additional length to each span of thread.
Now, slowly, carefully, submerge this threaded ring into still water.
As the ring is gently removed from its bath, a number of physical properties compete to shape the form of the threads.
    The structure of the wool ensures that each thread finds its way from one side of the ring to the other.
    The force of gravity tugs on the threads such that they sag in proportion to the amount of slack we allow.
    And the surface tension of the wet wool tends to bind nearby threads together.

Daniel Kohler and Rasa Navasaityte

As we remove the ring from its bath, these forces are resolved, and from their resolution emerges a surprising and useful property.
When viewed from above, the threads describe a set of minimal paths - that is, a configuration of lines that connect a given set of points in the most efficient means possible.
And so, our ring and thread machine may then be viewed as a simple, yet very efficient computing device, as it performed a useful calculation instantly, simultaneously, and without us needing to perform even simple arithmetic.

Centre for Information Technology and Architecture (CITA): Hybrid Tower

Before I go on - just one more that I'm sure you've heard.
This one begins with an axiom "as hangs the flexible line, so but inverted will stand the rigid arch"
These are the words of Robert Hooke, who discovered the relationship between a hanging chain, which forms a catenary in tension under its own weight, and an arch, which stands in compression.
We might be forgiven if we attributed this quote to Gaudi, who famously instrumentalized the principle of the hanging chain as a design method.

Jesse Louis-Rosenburg and Jessica Rosenkrantz: Floraform and Hyphae

Both of these - the wool thread minimal path calculator and the hanging chain weighted catenary calculator - are examples of what I would term a "material computation", wherein a natural or physical process is exploited in the service of calculating some useful property. 
Long before the advent of digital computation, material computers such as this had been put to work in experimental architectural practice as instruments of design - as machines for arriving at forms that the human imagination, on its own, could not.
They are also archetypes that have attained near mythological status within some cultures of architectural design, and have served as the inspiration for generations that followed.

Theodore Spyropoulos

A splintering of minor design movements identify the experiments of Gaudi and Otto as progenitors. These include: 
    generative design, parametric design, emergent design, bio-mimetic design
These design movements have since developed even more elaborate means for achieving similar ends,  means that include both digital and physical calculation.
I'm sure many of us in this room are familiar with the experiments shown on the screen behind me, and take inspiration in the complexity and beauty of the forms that result.
We should also recognize that there are likely many among us who find this sort of thing a silly way to go about design. Or, more bluntly, a way of avoiding design altogether.
In fact, there are many architects that totally reject form-finding as a valid approach to design. 

Jesse Louis-Rosenburg and Jessica Rosenkrantz

Critics sometimes characterize the contemporary design movements I mentioned above as "techno-rationalist", and find them to be at odds with the core approach long held by architects as central to what we do.
I describe what I mean by this in a section below, but for now I'd like to state that the motivation for this studio is that the critics of the application of form-finding in architectural design have a pretty good point. 
If we take seriously those that reject form-finding, I believe that those of us that choose to employ it will find entirely new approaches to the technique, and unexplored territories for creative application.

Zhu, Park, Isola, Efros: CycleGAN

When we factor into this mix new developments in machine learning, which echo form-finding experiments of the past while suggesting entirely new paradigms of design tools, we may find reason to believe that we are at the cusp of a whole new set of potentials in creative human-machine collaboration.

I'll take just a moment to explain why I see a correlation between the form-finding experiments of architecture's more distant and recent past, with a new algorithm that makes horses look like zebras.

Machine-Found Forms

Increasingly realistic synthetic faces generated by variations on Generative Adversarial Networks (GANs).
From left to right:
Goodfellow et al (2014), Radford et al (2015), Liu and Tuzel (2016), Karras et al (2017).
From General Framework for AI and Security Threats

The first thing to understand about the current state of AI is that techniques have gotten really good really fast.

These images are not photographic manipulations, but are completely artificial. They were produced by techniques developed by a branch of ML called "generative models". 

To train a generative model, a large amount of data is collected in some domain...
    (e.g., think millions of images, sentences, or sounds, etc.) 
...and a model is trained to generate data that resembles it.

The image on the right is not a real person, it is a synthetic image that resembles the data on which the model that created it was trained.

In this case, a dataset of photographs of celebrities.

Karras, et al: Progressive Growing of GANs for Improved Quality, Stability, and Variation

A similar technique was applied to produce the "morphing" images we see here... none of which are "real", or directly sampled from a photograph. These, too, are 100% synthetic images of celebrities, cats, bedrooms, and churches.

So, impressive as these uncanny fakes might be, what relevance might this technology hold for design in general, or for our understanding of form-finding in particular? To understand this, we must understand the nature of this new technology.

Machine learning has been defined as: learning through observation, in which patterns are mapped onto other patterns, without any intervening representations. What may be captured by the term "patterns" in this statement is remarkably broad.

Consider the following example, which begins to approach what we would consider a design tool.

Hesse: Edges to Cats
Here, an ML model has been trained to understand the transformation from a line-drawing of a cat to a photographic image of a cat. Once training is complete, this model will attempt to create a cat-like image of any given line drawing.

It is crucial to note the difference between the way a model such as this is trained, and how it is used. Just because this tool was seemingly made to create images of cats, it is more compellingly employed to make unexpected forms. Catopus, cat-car.

Google Magenta: Sketch RNN
Here, the drawings of an author are augmented with predictions of what is to come next. The model underlying this tool was trained using Google Quickdraw.

Google Magenta: Sketch RNN
The same model as in the previous slide, with this visualization showing many possible futures for the sketch. The model underlying this tool was trained using Google Quickdraw.

Such an approach would apply equally well to images of architecture.

Hesse: Regions to Facades
Here, an ML model has been trained to understand the transformation from a color-coded diagram of a building facade to a photographic representation of a building facade. Once training is complete, this model will attempt to create a facade image of any given line drawing.

Creative design tools have been fashioned using this same principle.

Note the similarity between these speculative design systems and those more traditionally understood as "form finding" approaches to design. In each, an author enters into a "conversation" with a metaphorical other.

Suggestive Drawing among Human and Artificial Intelligences
Here, an ML model has been trained to understand the transformation from line drawings to a whole range of objects: from flowers to patterned dresses. Deploying this model in the service of a creative design tool, Nono Martinez Alonso demonstrates the potential of computer-assisted drawing interfaces.

I have made some of my own experiments using the same underlying model, which is called Pix-To-Pix( ), and was developed right here at UC Berkeley.

Steinfeld: Depthmap to Death Valley
Here, using data from Google Street View, a model has been trained to understand the transformation from a "depthmap" and a photograph of a particular cityscape. Once training is complete, new synthetic photographic cityscapes may generated from arbitrary depthmaps.

Steinfeld: Depthmap to Delft
Here, using data from Google Street View, a model has been trained to understand the transformation from a "depthmap" and a photograph of a particular cityscape. Once training is complete, new synthetic photographic cityscapes may generated from arbitrary depthmaps.

The Work of the Seminar

The goal of this course is to assist students in the development of an architectural thesis proposal. While the theses we develop here are individual, we will each seek to situate our work in response to the larger theme that I've described here.

Although I have presented a number of examples drawn both from historical precedents and from some very recent developments in design technology, I would reiterate that all of these fall within the bounds of the metaphor I presented at the top of the presentation: that the artifacts of design thinking may themselves be seen to have a voice, and function as an active participant in the process of design.

This thesis section sees the term "form finding", along with the artifacts and design cultures related to it, in the broader context of human-machine collaboration, and seeks to update this project in the age of parametric modeling and machine learning.

While we all will embrace the idea
that the artifacts of design
may be seen to have a voice of their own,
students will individually choose their partners for conversation.

Regarding the artifacts of design as an "other" that is serviceable to the designers that wield them has functioned as a useful metaphor across a range of approaches and design design movements. All are welcome. Students will be free to sample and choose among a given set of possible projects, a set that will accommodate a range of design positions and technical requirements.

To briefly mention some approaches that would be considered "fair game":

Some have sought to literally delegate certain carefully-chosen aspects of the design process to others, thereby expanding the agency of certain stakeholders and altering the social dynamics of design.

Jose Sanchez, Blockhood (left), Herman Hertzberger, Central Beheer Office Building (right)

Some have sought to externalize and make explicit the values and objectives of a design, thereby cleaving the design process into those decisions that concern the character of the tool and those decisions made while wielding it.

Marc Fornes, Non-Lin / Lin (left), Frei Otto, Munich Olympic Stadium (right)

Some of these imagined "others" were created to be stoic and rational, serving to impose constraints upon the imagination of those that wield them, while others were specifically crafted to evade rationality, functioning as a creative prompt that stimulates a designer to transcend the constraints of internalized convention.

Norman Foster, London City Hall (left), Brian Eno and Peter Schmidt, Oblique Strategies (right)

The seminar will explore these issues through the preparation of a thesis book and digital presentation, as well as through the development of four devices:

An Experimental Physical Model

Physical Experiment -> Building; Francois Sabbio
Physical Experiment -> Building; Julia Spackman and Alex Still

An Experimental Parametric Model

A Video of a Kangaroo Model I Downloaded at Random from YouTube.

A Board Game

Board Game -> Building; Dongil Kim, 2014
Board Game -> Building; Philip Goolkasian, 2014

A Generative Adversarial Network

Synthetic Rose Petals; Sarah Meyohas, 2018

House forms generated by a GAN

Background: The Composing vs Finding of Form

Architecture has historically been positioned as a synthetic practice, one that centers on the the ability of the designer to make conceptual leaps that connect disparate sets of demands, and unite seemingly unrelated orders.

The representations we employ as designers hold significance only insofar as they help us to find something new.

In contrast, the form-finding model of design has widely seen as primarily analytical. It hinges not on any interior capacity of an individual designer, but on the efficacy of a carefully-crafted design experiment to appropriately model some aspect of a useful architectural reality.

The hanging chain is a valuable partner because it produces structurally useful forms.

The soap bubble is a useful partner because it describes minimal surfaces, which are useful for designing tents.

Note the similarity between the way that each of these models value representation, each employing the metaphor of an 'other' that assists us in our design endeavours.

We can see this metaphor at work in that way that compositionally motivated architects often personify the artifacts of design production, speaking of what a drawing is "saying", or noting that a model "wants to be" some other way.

While the proponents of form-finding in design might not characterize it as such, a similar authority is bestowed upon the physical artifacts they employ.

In each case, the practitioners of these methods find it useful to imagine that the artifacts of design thinking themselves have a voice, and function as an active participant in the process of design.

Understanding this role of the artifacts of design as an "other", one that is serviceable to the designers that wield them, offers us the key to understanding the challenges associated with integrating form-finding techniques in creative architectural design, and the key to the connection between classical form-finding and contemporary machine learning.

To be effective, compositional design relies upon a specific set of qualities in the artifacts it employs.

Drawings and models are most useful when they exhibit just the right amount of ambiguity: they should be concrete enough as to suggest a new order that we might not have otherwise considered, and yet flexible enough as to allow us to project new orders into them as we see fit.

Compositional designers seek in their representations good partners for conversation - affable, engaging, and generous - qualities that are too often overlooked by the more stoic cultures of practice engaged in form-finding.

The experiments associated with form-finding design techniques are too often regarded as necessarily generative of unmodifiable answers to design questions.

When taken is such a rigid way, the failure of the form-finding models we see here becomes clear: they are ineffective partners in the conversation required for productive design thinking.

Thank you for reading.

For more information, please see the draft course syllabus on Google Docs.