This post was originally published by Jonathan Follett at Towards Data Science
Algorithmic Brainstorming for Creative Work
By Dirk Knemeyer and Jonathan Follett
Architecture, like many creative professions, spans both the digital and physical world. As technologies such as 3D printing move the discipline forward in remaking the built environment, AI and generative design are having an impact on architecture from a planning and design perspective, where the profession is largely digital and computational. We spoke with Lilli Smith, Senior Product Manager AEC Generative Design at Autodesk, a practitioner in the field of architecture for more than 20 years — the last 18 of which have been making the software that architects use to design their creations.
In architecture, art, and other creative fields, generative design is a methodology that automates the creation of design options that may balance a variety of competing goals. The latest wave of generative design is driven by artificial intelligence. “This is often the case in architectural design problems,” Smith says. “There’s not a single objective that you’re trying to go after — there’s no one answer. But, there’s a lot of different objectives that come into making an architectural project. So, for example, in an urban design workflow, having a nice amount of open space and good views from your property might be important. But return on investment and value generated through more rentable area[s] might also be important.”
Generative design can help automate the creation of options, which satisfy a variety of goals that the designer wants to encode into the system. Generative design can also be an exploratory tool to open up a designer’s thinking — not necessarily solving the problem or providing one right answer.
“Using algorithms to guide the creation of these design options, the computer will give you … the best options that it can, given the goals and the size of the exploration also that you have specified,” says Smith. “Then you can have a discussion with your stakeholders about which goals may actually be more important or which designs … the stakeholder may favor for other reasons. It’s a great way to have conversations with people.”
“When there are several inputs to the design, it becomes really hard for the human mind to keep track of all the combinations of those inputs,” says Smith. “The computer can actually surprise you with combinations of different inputs that you might not have thought about before …”
The AI can quickly produce a high volume of concepts attuned to the requirements of the project. That concept generation, or what we might have used to think of as “brainstorming”, at scale, by the machine is affordable and more comprehensive than what a human could do in the same budget of time and money. This is particularly important in the context of architecture where, as Smith illustrates, the requirements are complex, extensive, and can often be competing. “So, the computer isn’t really coming up with anything on its own,” says Smith. “[It’s] just having an automated way to produce all of these combinations that humans aren’t good at, but computers are really good at. And then being able to rank and sort them, or search through them, given the different metrics that you’ve also included, can sometimes even yield surprising and delightful results that you didn’t expect.”
A Brief History of Generative Design
Even though we think of terms and applications like “generative” as cutting edge, programmatic artists have been using this creative method for over 50 years. “Generative design is really not new,” says Smith. “There’s a long history of generative art.” In the 1960s, the pioneering Hungarian computer artist Vera Molnár worked in the early programming language Fortran to generate images examining theme, variation, automated generation, and display of options in her work. Digital artist Manfred Mohr, another pioneer of algorithmic art, generated variations of 3D geometry in the 1960s and 1970s. And, in their 1972 article, “Shape Grammars and the Generative Specification of Painting and Sculpture”, design technologists George Stiny and James Gips created a manual encoding of design systems, describing a production system for generating shapes. “It’s really a direct line of thinking for a lot of what goes on in parametric design and also underpinnings of tools like CityEngine, which is a tool by Esri for urban planning based on GIS data, and other tools, as well,” says Smith. “There are papers in the 1970s that demonstrate how people were really thinking about, in architecture and specifically in hospital floor plans, how to lay out those floor plans algorithmically,” says Smith. “So, these ideas have been around for a long time in architecture. They’ve also been around in engineering.”
In a more recent example of AI-driven generative approaches, NASA used programmatic design to create configurations for its satellite antennas. Jason Lohn, the leader of this project at NASA Ames Research Center, described the algorithmic approach to design in a feature article published by the agency in 2006:
“The AI software examined millions of potential antenna designs before settling on a final one,” said Lohn. The software did this much faster than any human being could do so under the same circumstances, according to Lohn. “Through a process patterned after Darwin’s ‘survival of the fittest,’ the strongest designs survive and the less capable do not.”
“We told the computer program what performance the antenna should have, and the computer simulated evolution, keeping the best antenna designs that approached what we asked for. Eventually, it zeroed in on something that met the desired specifications for the mission,” Lohn said.
“They searched through tens of thousands of these different antennas,” says Smith. “And they were able to come up with a design that was never something that they would have come up with in their own minds [through] their own traditional design process. But the final antenna design performed more than 90% more efficiently over the traditional antenna.” In the NASA example, we can see how AI can inform and enhance creative decisions by introducing superior design options that humans may not even initially consider, with surprisingly effective results. The design output is better because both human and machine are doing what they do best.
Computational Design Approaches in Practice
Inspired in part by generative design’s storied history, Autodesk research has long been interested in using generative workflows. “A couple of years ago, Autodesk bought an architectural firm called The Living out of New York City,” says Smith. “[An] amazing collection of talented designers who also have computational brains. They are really good coders and very good designers. And, they’ve worked on a number of projects involving generative design pursuits.”
“They worked with Airbus on a partition. [It’s] one component of a plan which might not seem super important. But, there’s a lot of different functionality that has to go into the thinking about that panel. They have to be able to remove certain sections of it to accommodate emergency stretchers going into the plane,” says Smith, “… and of course rigorous safety standards and crash testing standards.”
“They were able to use this process of studying flexible models and looking at how a material could be taken out, but you still have the structural integrity of this panel. And also use 3D printing technology to get the total weight of each partition down by about 45% or about 30 kilograms,” says Smith. “And this might seem minor, in the context of a whole plane that’s really big and fully weighted. But, they do estimate that these weight savings across the current generation of these A320 series of planes could result in the savings of almost half a million metric tons of CO2 every year.”
In order to test their applications of and theories about generative design, Autodesk used the method in the design of their new Autodesk Toronto facility — referred to as their new center for emerging technologies. “They wanted to use generative techniques in order to design the new office space. So, they started out by collecting a bunch of data to guide the design and help them to create goals for what was their ideal office,” says Smith. “So they ask people questions like: ‘Who would you like to sit by?’, ‘How much distraction can you take in your day?’, ‘Do you like a lot of daylight near your desk or do you need a darker environment because you’re using screens and you don’t like a lot of glare’”. After conducting extensive internal research, the team developed six goals that they used to evaluate their designs. “They surveyed existing employees and then they boiled all that data down to these six goals that were about adjacency, distraction, things like that,” says Smith. The team created a flexible model. “They had an idea about how this design was going to work and had an idea about how it could flex and vary,” says Smith. They wanted to study … the various implications of where they put groups of desks and where they put groups of amenities, meaning conference rooms, phone booths, more private spaces. So, they developed flexible models and ways to measure the success of those designs and use the computer to help guide them to solutions that satisfied those goals.”
AI and the Future of Architecture
“A lot of people that we talk to see the possibility of the computer augmenting their design efforts,” says Smith. “They see that humans are still going to be critical in these design efforts because they’re going to be setting up the problems, deciding what kinds of problems to solve, using machines to help them to do a better job.”
“I think we’ve got a long way to go before the machines are writing all the software or doing design for us. Maybe it will happen someday but I think we have a long way to go before the bots take over,” says Smith. Our computer overlords remain a long way off. The machines are going to be helping us to make things, not removing us from the equation. They will remove the more physical, manual, mundane, and tedious parts. They will not be doing the core problem solving, particularly up front in the research and requirements but even, for the time being, in the core solution articulation once a generative approach has determined the correct basic direction.
“I think generative design is a different way of thinking for designers. It’s different to think about designing a system for design instead of several one-off designs…. It’s different to think about how you can make a whole design system that’s going to flex for you,” says Smith. “It also requires computational design skills, so that you can use the computer to augment your powers. We’ve tried to make it easier for architects and engineers by making these visual scripting environments where you can write code by just putting together these different nodes or different functionality. And we also work on developing nodes that people can share. And also developing a community of people who want to share code with each other.”
Enabling creators via a visual interface that precludes the need to learn programming or otherwise tangle with unsophisticated UI is important to unlocking the potential of artificial intelligence in a creative context. The approach of offering visual environments of use, instead of expecting non-programmers to use programming to best use next generation tools, is a common theme when it comes to emerging technologies like AI. Contrary to the popular buzz around the importance of everyone learning programming, the reality is that companies that create the tools for creatives are proactively trying to make these tools so they can be fully exploited by the non-programmers among us. Indeed, while today there is a good argument for many creatives to learn some degree of programming to best leverage the early and nascent tools at our disposal, those are skills that might build our long-term holistic knowledge base, but will only be practically applicable in our work for a shorter time horizon, during this transition period when computational creativity is in the early days.
Using AI-assisted automated methods to better design buildings may very well be a future imperative. “By the year 2050, there are going to be 10 billion people on earth. And, if you do the math, we’re going to need to build about 13,000 buildings a day to accommodate all those people,” says Smith. “In order to build all these buildings without totally destroying the earth, we’re going to have to develop better methods to design and construct [them]… Generative design is one tool in this effort. I think we’re going to have to drastically change a lot of how we’re getting things done to accommodate all of these people on earth.”
Creative Next is a podcast exploring the impact of AI-driven automation on the lives of creative workers, people like writers, researchers, artists, designers, engineers, and entrepreneurs. This article accompanies Season 3, Episode 6 — Architecture and Generative Design.
This post was originally published by Jonathan Follett at Towards Data Science