Archllucinations: Inhabiting machine dreams
How can we teach architecture to an artificial intelligence?
Everything that has a name, exists
It can be said quite correctly that, since man is able to abstract the spaces in which he lives, the history of architecture goes hand in hand with the history of its representation. From the two-dimensional abstract drawings, through the discovery of the laws that govern perspective and design, to modern computer work techniques, architectural representation is a preliminary step to architectural materialization.
This statement demonstrates the limitations that the architectural design has by default: it will be based on the knowledge, experience and resources of the designer. As the saying goes, everything that has a name exists; therefore, what has no name does not exist. If everything depends on that experience and knowledge, are we not missing out on a great field of opportunities with everything that is not written, that is not thought beforehand? What about all those spaces that have not been imagined due to the limits of the mind that designs them?
The question to ask would be, can Artificial Intelligence help to bridge that gap?
Of course, before that, we have to teach AI what architecture is.
How can an artificial intelligence learn architecture?
The extensive and methodical work that Stanislas Chaillou does in his “ArchiGAN” project worked for us as inspiration, and made us realize the importance of working in 3 dimensions. The choice he makes of the 2D dataset seemed too deterministic, since it continues working on the plane representation of architecture, and not on its spatiality.
So, if we train with a 3-dimensional dataset, the question is clear, how does the machine understand the difference between a three-dimensional space and an architectural space?
Given this question, we have to think about how we, the people, turn the physical space into an architectural one: we inhabit it, and not only in a static way, but we also go through it.
The routes are chosen when considering them a transversal variable that allows to perceive the geometric proportions of each space, its sequential relationship, and adds the essential part of the user experience while inhabiting it.
Against all the qualities that influence the perception of a space, such as lighting, texture, temperature, smell …, what is perceived by sight can be the most influential, and therefore, the chosen one for this first test.
So, let’s put our virtual machine to go through space and train with what it sees.
Let’s walk: Dataset generation
The model we have developed has been trained with a dataset of three-dimensional models of “master” works of architecture, belonging to the Swiss architect Charles-Édouard Jeanneret, known as Le Corbusier, with the intention of checking if the AI is capable of abstracting characteristics like the style.
The process begins by drawing a route in each building, using a 3-dimensional curve with an average observer height of 1.65m.
Then, a script has been programmed, using Grasshopper and Python, which generates a virtual machine that runs through each curve, projecting a sphere of visual rays in all directions every 5cm.
The isovists collide with the geometry giving rise to intersection points, which are referenced by 3-dimensional coordinates whose origin is always the head of the observer.
The vectors that generate these intersections have the same directions at all times, so that the data with which we train is the module of each one of them — its intersection distance -, which allows its simplification when training the AI.
In addition to these distances, point coordinate information (Xt, Yt, Zt) is also stored in the path, the tangent at that point (αt, βt), and the position (t) remapped of the point as a function of the curve.
In this way we are able to abstract spatial paths in data with which we can train mathematically. Once all the points have been collected, they are organized in 2 formats for the training:
The first format is to develop the distance information of each point in an array (A1260). Thus, at point t of the curve traveled, the data structure would be:
Pt = ((Xt, Yt, Zt), (αt, βt), Tt, At (1260))
The second format, all visuals of each sphere are displayed in a matrix M that would coincide with a mercator projection. This option is of great interest since it generates more relationship between the points that are close in space, which will be very useful later. Its structure would be:
Pt = ((Xt, Yt, Zt), (αt, βt), tt, Mt (21.60))
AI model selection
Once we have the dataset, that is, once the AI has “traveled” all these buildings, will it be able to generate architecture?
The logical choice of a DCGAN (Deep Convolutional Generative Adversarial Network), which faces a generator to a discriminator in an unsupervised way, allows the AI to decide which are the important patterns and relationships at the time of the generation of those spaces .
The neural network has been trained independently with the two morphologies of dataset, array and mercator. To enter the information in mercator mode, an encoder has been used prior to the introduction into the network, and a subsequent decoder.
Dreaming Architecture
If AI has learned architecture through it, it has to walk to imagine a new space. So, the input you need is a new route, a 3D curve, which allows you to suggest spaces around it.
These new spaces, are still on the edge of the dream, hallucination, and allows us to intuit proportions, gaps and a concatenation of different scenarios. That is, the AI ”imagines” and proposes architecture as it progresses, though the person, the human is always at the core of the process.
Will Archllucinations serve as a proactive design tool in the future? One that helps the generation of a more humane, higher quality architecture?
Impact
The project is currently at a conceptual point. The idea is to continue developing and including other key points when analyzing architecture. That is, adding in addition to geometric information, qualitative information.
This information could be lighting, material, temperature, energy efficiency, orientation, function, adjacent spaces, as well as proportions depending on the program or the number of people that have to inhabit each space.
Faced with the inclusion of computer science in the design of architecture in a deterministic way, as it has been until now, Archllucinations would become a proactive tool, easily implementable in the current BIM methodology (both in data collection and in the propositive part)
Going back to the initial analogy, if “everything is written”, if “everything exists”, if rather than depending on his own acquired knowledge the designer had a series of architectural quality options to be inspired with, would we not be stimulating good architecture?
Conclusion
It is essential that we have tools that help us to democratize good architecture. It affects us all from different because we live in it; we live surrounded by it, and the fact that up until now it is has only been overseen by people who think about the economic benefit and not in its more social, artistic and psychological component, significantly impoverishes our lives, with the unawareness of most people.
The city, as the setting for our social, cultural and psychological activity, needs to move away from purely economic models and reconnect with the citizen.
Archllucinations wants to demonstrate that we can approach a matter of vital importance for society such as architecture in new ways, taking advantage of the emergence of AI. We want to focus on improving the quality of the spaces in which we live. In the immediate future ⅔ of the population will live in cities, them being the seed of this artistic process.
Saturdays.AI
No one:
Absolutely nobody:
Architects: could AI make buildings?
Looking for a solution to this problem, we arrived at AI Saturdays, a non-profit group that helps introduce people from different fields into the world of artificial intelligence, encouraging them to find applications of these new technologies in their fields of knowledge.
Teaching is purely practical through exercises ranging from the initial application of linear regressions and decision trees, to the use of neural networks, ending this process with a guided project of 2 months.
Project developed between:
Francisco Jose Rueda