Doom or Bloom: What Will Artificial Intelligence Mean for Architecture?
Is Artificial Intelligence (AI) the doom of the architecture profession and design services (as some warn) or a way to improve the overall design quality of the built environment, expanding and extending design services in ways yet to be explored? I sat down with my University of Hartford colleague Imdat As. Dr. As is an architect with an expertise in digital design who is an assistant professor of architecture and the co-founder of Arcbazar.com, a crowd-sourced design site. His current research on AI and its impact on architectural design and practice is funded by the US Department of Defense. Recently we sat down and talked about how this emerging technology might change design and practice as we now know it—and if so, would that be such a bad thing?
MJC: Michael J. Crosbie
IA: Imdat As
To start off, can you tell us a bit about your research into AI as it pertains to architecture?
Before I get into AI, let me give you a wider context. Over the past five years I’ve been working on “crowd-sourcing” architectural design. I co-founded Arcbazar.com, an online competition platform, which crowd-sources design challenges from all over the world. So far it has hosted thousands of design competitions. The majority is for residential projects—about 55 percent of the competitions—and the rest is divided among interior, landscape, and commercial design projects. In the process, we accrued vast amounts of design data. In 2017, we (Raytheon and the University of Hartford) received a grant from DARPA, the Defense Advanced Research Projects Agency of the US Department of Defense, through its Disruptioneering program, and we used Arcbazar’s design data to study the potentials of AI in design. As part of our research we developed a function-driven “deep-learning” system, a popular form of machine learning, to explore the potential of generating conceptual designs with AI.
Artificial Intelligence is a term that describes a great many developments in our world right now. Can you give us a clear definition of what AI is?
Artificial intelligence is simply when machines mimic “cognitive functions” we normally see in humans, such as “learning” and “problem solving.” Popular AI applications are, understanding speech (Siri or Alexa, for example), learning to play games such as chess, or being able to drive a car. Simply put, AI is intelligence shown by machines in contrast to natural intelligence we see in humans.
What’s the pace of how AI is developing, the velocity of this change?
AI has actually been around for a while. The term Artificial Intelligence was coined at a conference at Dartmouth College in 1956 organized by Marvin Minsky—the founder of MIT’s AI lab. Since then there’ve been highs and lows in AI research (the slower pace of research was related to limited computer power, data availability, funding). Since 2011, however, there have been major developments in machine learning, in particular due to research in deep neural networks, or DNNs. DNNs are just one area of research in AI; there are others, such as evolutionary computing, expert systems, genetic algorithms, fuzzy logic, for example.
DNNs loosely mimic the inner workings of the human brain. They have various layers with neurons, an input layer (where data are fed into), hidden layers (which process data, although we’re not sure how they actually work), and an output layer (which produces the result). One can manipulate the weights of neurons, the number of hidden layers, and so forth, thereby iteratively improving the performance of the system. One of the early uses of DNNs was labeling images. A DNN system, for example, was trained with millions of images of, say, dogs. The DNN “learned” what a dog looks like through a discovered internal representation, and then could correctly identify a dog in any new image inputted. With the Internet coming of age and the availability of large amounts of data, a new golden era of AI research is emerging. Most of the papers on generative adversarial neural networks, known as GANs—which are a new variation of DNNs—have emerged as recently as 2017. The pace of research and development in this field is mind boggling. Consider this: 8 out of 10 new startups claim to deal with a form of AI: for healthcare, for finance, and so on. The reported world market for AI hardware, software, and products reached about $8 billion this year, and is projected to reach $90 billion in seven years.
So far you’ve done a great deal of collaborative research with computer scientists on architecture and AI. What are some of the critical things you’re learning?
In real estate, you know, it’s all about location, location, location. In AI it’s all about data, data, data. Our first experiment was to train a DNN with home designs, so it could evaluate homes based on particular functions, such as the performance of living spaces, or kitchen/dining related activities. We manually scored the quality of designs on a scale of 1 to 100, and then asked the DNN to evaluate new designs on its own. Amazingly, the DNN scores were very close to the scores we had originally given. To give you an example, we scored a kitchen/dining space arrangement, say, at 51, and the DNN scored it at 51.2.
Once the DNN was able to evaluate designs with such accuracy, we asked it to identify essential building blocks from within a larger home design, corresponding to a particular target function. The idea was to re-compose the homes’ high-performing smaller building blocks into new assemblies that could be used in new home designs, similar to creating “Frankenstein designs” out of bits and pieces from existing data. What’s important in such research is not only the amount of relevant design data but also the format—is it an image, a graph. All data need to be uniform and machine-readable.
What are the most direct applications of AI to architecture on the horizon?
In the short term I think we’ll see AI-driven CAD-software that can assist architects in the early stages of design development. The software tools we use now, like Revit, are actually not very good in early conceptual phases of a project.
Conceptual design requires a lot of exploration, testing several ideas at the same time. The best way to do conceptual design is still with pencil and paper. However, with new AI-driven software, a designer might provide a host of constraints, say, a chair made out of a particular material that can hold 300 pounds. The software could generate hundreds of optimal chair variations that the designer could choose from and develop further.
In the long run I can see AI-driven systems performing more complex tasks. Like, designing a home for two adults (one who uses a wheelchair) with two kids of certain ages and genders, a dog, on a particular lot. The system could pull all zoning data, building codes, and disabled design data, and generate design variations that also follow a certain design vocabulary and offer options—perhaps directly to the client, who picks the most responsive design.
What do you see as the potential benefits of AI to architectural practice?
Deep learning is a very powerful analysis, identification, detection, and classification tool. Where data are available, AI can help architects with lots of analytical tasks, to have a better understanding of context, circulation patterns, spatial and material performance. This is true for not only the quantifiable and obvious characteristics of design, but also the more non-quantifiable ones, like maybe how a space makes you feel. We usually can’t easily quantify these qualities and they often get neglected.
What do you see as the potential benefits of AI to architectural design—what will be the impact of “deep learning” by machines on generating designs?
Deep-learning machines could decipher patterns in architectural design that architects have intuitively or intentionally created over the years. Think of Christopher Alexander’s A Pattern Language, which compiled various patterns in the way we design residential architecture. AI could potentially expand such patterns to include not only functional and programmatic concerns, but also socio-economic, ideological, geographic, climatic, or other patterns that shape the built environment. With such a resource, AI-software could assemble the best patterns for a given problem into new compositions. Design decisions can become much smarter, which might lead to new hybrid vocabularies, and novel dynamic, adaptive, and synthesized compositions we never thought of.
AI can “learn” to design according to functional aspects of architecture; but what about those intangible qualities of architecture—human fascination, amusement, even spirituality? Can AI incorporate these qualities into design?
Deep learning is all data driven, and any aspect of design, even intangible ones, can be analyzed—given there are data. For example, you could train a DNN about what makes a person of a certain culture perceive an architectural space as “spiritual.” You might collect data on such spaces and their ability to provoke spiritual feelings (based on perceptions recorded from visitors of various sacred spaces). The DNN could decipher major patterns that it identifies as essential to create a spiritual space. Looking at thousands of examples, the DNN might discover that such human perceptions of spirituality occur because spaces have certain proportions, lighting, height, scents, or aural qualities. Some of these might be obvious—the result that the architect intended. But others might be latent, that we never thought of. Such work is really based on the availability of data, and the more the better.
What are the possible threats of AI to current architectural design and practice? How could it radically change the way architecture is created?
I don’t really see AI as a threat. As a technology that can eventually help designers and clients drastically improve their built environment, I think with emerging AI tools designers will benefit from a swarm of design ideas very early on—one could compare it to crowdsourcing designs online but with no human designers. In addition, all the grunt work we do as architects could potentially be automated. We then could reach a much wider audience, and provide access to quality design to a broader portion of the public.
How do you think AI could change the role of “starchitects”? Will AI be able to do better Gehrys than Gehry?
There are actually style-transfer applications on 2D images, text, or audio. For example, “Neural Doodle” by Alex Champandard translates any rough doodle into a painting by a famous artist. At some point this will be achieved in 3D as well. It could be a hybrid system where a graph-based DNN like the one we worked on is combined with an image-based DNN that identifies a personal design vocabulary or style and applies it to a new composition. Such a development has the potential to impact practice. This is wild speculation, but perhaps architectural practice could be modeled on the music industry, where, for example, Frank Gehry develops a “style” and whoever uses his language through an AI-driven system pays him a royalty. Gehry in that way might “design” millions of structures around the world. We could even have architects who long ago ended their careers “designing” again via AI.
Some observers warn that AI might be the end of human civilization, as machines take over decision making from the planet’s “inferior beings” (that’s us). Do you see a similar potential in architecture?
There is a project by MX3D, a robotics start-up company, to 3D print a bridge on one of Amsterdam’s canals without physical human intervention. If in the future 3D printing could happen autonomously, we might have intelligent, self-constructing 3D printers that build, re-adjust/re-evaluate, build again, based on climatic conditions, circulation methods and paths, safety concerns, and so on. It would be interesting to see what type of architectural spaces, urban layouts they might develop.
I don’t predict a “Westworld” scenario with machines taking over humanity, at least not for the foreseeable future. I actually think the opposite. With the problems we are facing on earth—like climate change, overcrowding, poverty—AI is perhaps an essential technology that will allow a major leap in human history to organize, produce, and improve. The problems we are facing are very complex, and will be difficult to solve without AI intervention.
Five years from right now, how do you think architectural design and practice will be different because of AI?
We always talk about the Renaissance master builder as the archetypal designer, who was not only an architect, but also a builder, an engineer, and was able to marry fields of the practice that we now separate. The power of building information modeling tools like Revit or ArchiCAD was that they somehow brought back the power of the master builder, because they enabled a single creator to not only design the edifice, but also model structural efficiency, simulate energy consumption, and other factors. They empowered the architect, so it is not surprising that 72 percent of all architectural offices in the US are one-to-two-person offices. Perhaps this trend will be further enhanced with AI coming into play. Not in five years, but perhaps further down the road, AI-driven design systems could be used directly by clients. I think this will help the 90 percent of construction work currently done without architects. These AI-driven design tools for the non-professional would be created by architects, allowing them to extend or expand their design knowledge and influence to areas of the built environment they currently don’t have access. It would be a blooming of quality design accessible through AI-driven design software. One way or another AI will have a deep impact on the way we conceive, represent, and develop architecture and shape our built environment. I think this is truly a turning point in architectural history.
Featured image courtesy of the author.