Vahid Moosavi
Image @ Vahid Moosavi

Data-Driven Modelling for Architecture and Design: A Conversation With Vahid Moosavi

Share on facebook
Share on twitter
Share on linkedin
Share on telegram
Share on pinterest
Share on whatsapp
Machine learning and data-driven modelling offer a radically different way of looking at the world than traditional domain-focused and theory-based approaches. Machine learning algorithms have created a wave of computer vision and natural language processing applications for a wide range of fields. These same algorithms are now changing how we think about and solve urbanism, architecture, and design problems.

In recent years significant advances in reinforcement learning, generative models, and new datasets, such as the Fusion 360 Gallery dataset, have created new opportunities for content creation, design exploration, and new tools to aid designers. A critical question that remains open is how we should integrate these new methods and approaches in the design process. In other words, what will the role of the designer be as these new technologies evolve? How should humans and machines interact to solve design problems?

We talked to Vahid Moosavi, a former senior researcher and lecturer at CAAD, Institute of Technology and Architecture (ITA) at ETH Zurich. Moosavi has been working on data-driven approaches for architectural design and developing tools and methods that integrate human and machine intelligence to solve design and engineering problems.

Moosavi is a systems engineer who has had a nomadic journey that led him to work in architecture. He had worked as an IT consultant in various sectors, including manufacturing, energy, and health, before joining the CAAD, Institute of Technology and Architecture (ITA), at ETH Zurich. At CAAD, he focused his time on research and practice of Machine Learning and Data-Driven Modelling in the context of architectural design and spatial systems. Currently, Moosavi works as a senior data scientist at Swiss Re, building data-driven risk management products.
Learning Physics: Fast and scalable urban flood risk estimation, Learning to emulate slow physics based simulation engine. Image @ Vahid Moosavi

What initially interested you in applying ML and Data-Driven Modelling in the field of architectural design?

I have a diverse academic background, and over the last 18 years, I have been on a nomadic journey traveling across different fields. I initially studied systems engineering, including probability theory, statistics, mathematical modeling, and optimization courses. For a few years, I worked as an enterprise system designer and IT consultant in various industries, such as manufacturing, energy, and health sectors. In 2010, I gradually realized that most systems engineering education around the topics of “systems analysis” and “system synthesis or design” seemed to be very individualistic and based on rational methods. Usually, those design methods are mechanical and far from the real politics happening among the stakeholders. One could see that in the end, the success of a design project was dependent on the competency of the so-called “system architects”. This drew my attention to the field of architecture, from where many of the terms in systems engineering were borrowed.

Consequently, in 2011 I entered the architecture domain, and I started my Ph.D. at the chair for Computer-Aided Architectural Design (CAAD) at ETH Zurich. Since 2011 I have been focused on the research and practice of machine learning and data-driven modeling in the context of architectural design and spatial systems. I focused on theoretical research and techniques and was also involved in academic and business-oriented AI applications and startups. As a result, I have been closely working with many architects, structural designers, urban planners, and real estate professionals in Singapore (Future Cities Lab), Switzerland (ETH Zurich), China (South East University, Nanjing), and recently in the UK.
Machine learning and structural design. Image @ Vahid Moosavi

What opportunities have you identified in using data and AI in the design process?

To identify the opportunities, we need to clarify what we mean by AI and design. For me, AI (machine learning and data-driven modeling) is a generic framework that provides the possibility to learn reasonable answers to any question/problem we might have, provided that these problems are generic enough and that one can collect enough data around those problems. If these problems are not generic and repeatable, machine learning is not going to work. Classically, experts are those who know the answers to domain-specific questions, and in my opinion, AI is turning this upside down. Fields that are about problem-solving with domain expertise will be automated in large parts.

While AI is good at learning answers, I don’t believe that AI can come up with “good questions” or to decide where to pay more attention (Even though the recent breakthroughs in natural language understanding are precisely based on the so-called “Attention mechanisms”). Therefore, if we define Human Intelligence (HI) as the ability to develop the right questions, and Artificial Intelligence (AI) as the ability to learn the correct answers, HI and AI can play very nicely together, where the human actor is thinking about the question or the context. The machine takes care of those tedious tasks (e.g., searching in a database of millions of images) or sometimes tasks impossible for HI (e.g., learning a nonlinear function out of a large high dimensional data set). Therefore, I think AI will be massively implemented in applications with clear questions that are repeatable, standardized, and easy to automate. On the other hand, those applications which need more creativity (thinking out of the data!) and continually changing the questions and reframing the problem, or those that rely on strong storytelling skills will remain in the realm of human intelligence.

Now the question is, what do we mean by design? For our discussion here, we can imagine a spectrum where, in one extreme is more about creativity, art, and problematization. On the other extreme, it is mainly about scalability, repeatability, and problem-solving. If we think of engineering design and optimization, where we deal with more standardized products, like in the auto industry and other consumer products, there are plenty of opportunities.

Technically speaking, the first opportunities are definitely in surrogate modeling in design optimization, which is about speeding up the physics-based simulations such as Finite Element Analysis (FEA) in structural and mechanical systems or other Computational Fluid Dynamics (CFD) simulations such as water and energy in buildings or cities.

The second type of application is at the intersection of machine learning and generative tools in structural design. Here, we can use machine learning to learn the complex nature of these generative techniques. For example, a method called Force Density Method (FDM) was developed in the 1970s. Given a particular graph and the loading conditions, it solves a linear system of equations and gives us a geometric form under forces’ equilibrium. This is great as it is very fast, but now the question for the human designer is which system FDM should solve? FDM generates a new geometry and a new form every time you change the topology and other structural parameters such as the force densities on the input graph. Unfortunately, it turns out that these systems are chaotic and highly nonlinear and can quickly go beyond the limits of interactive systems. But now, since these generative methods are very fast in generating large datasets for specific design briefs, we have an opportunity to develop data-driven systems for the interactions between the human designer and the machine to explore the design space and navigate in the space geometric possibilities. Here, the machine acts like a personalized assistant that learns the human designer’s preferences, explores the design space, suggests new options, and gets feedback from the designer.

Another type of application of AI in design is in the field of building systems design and smart homes, which depends on the behavior of the people who will live in a specific building and a particular climate or region. At the moment, many design decisions about these building systems and how they should operate are made in the early stages of the design process, with a lot of assumptions about “normal” human behaviors. These assumptions can be questionable and falsified. Here, machine learning concepts such as Reinforcement Learning (RL) and interactive learning frameworks, together with IoT technologies, promise new building operation paradigms.

Another type of AI applications is emerging at a larger scale in cities. The ever-growing pervasive sensing and diverse urban data streams all over the world have introduced new planetary scale “comparative approaches” where cities can learn from each other. An urban designer can identify complex and high dimensional patterns in a specific city while having a detailed view of all the similar cases worldwide. I believe that these types of applications can be used in real estate applications, which are highly scalable compared to one-of-a-kind architectural design projects.

Finally, we can think about those (architectural) design projects, which are not really generic, repeatable, and scalable. In a way, these projects are interesting because they are “outliers”. For example, we can think of a famous building compared to the majority of buildings in a city. These rare design projects rely on “mastership,” and are heavily dependent on human intelligence, and I am not sure how one can use AI to generate a masterpiece. Because what AI is good at is nonlinear interpolation in the experienced world (data), while the next masterpiece does not exist yet. If it exists, it does not have a lot of similar cases in our datasets. Therefore, I am not personally interested in those applications of AI using state of the art in deep learning such as (GANs) to generate art. I think true art means being outside of (or against) a system, while AI is how to behave within the system and learn to comply with the rules.
We should think of AI as a “language”. While computer scientists are more focused on how to develop its underlying mechanics, the architects and designers should become fluent in this language, and beyond “writing” they should think about “what to write?”

What are the biggest challenges right now in using data and AI as part of the design process?

Overall, computer science, including AI, is a new field that is evolving quickly with new concepts introduced frequently. This brings opportunities and challenges. Whenever there is a hype about new technology, we should be conscious that many hot topics will not survive in the long term. Therefore, we need to be patient, not react quickly but think more about the philosophical implications of AI in design. Some of the main challenges that are not new and have happened over the last 70 years in computational design are that designers and architects have played passive users against the hot topics from computer science in different periods.

I think AI is hyped for a good reason, but this does not mean that the designers and architects should look for quick applications. We should find ways to act proactively, and this requires data literacy, critical studies of AI, and basic computational knowledge. Without these efforts, we just need to sit and wait for the next wave of hot topics, popularized by social media companies like Google and Facebook.

In addition to this general challenge, I think one of the primary limits of AI in design is the nature of architectural design projects, which are constrained to their physicality. For example, think about a building as a designed physical space, which is going to stay the same for a long time in comparison to an online market place, where a computer-agent continually learns how to recommend new products to the users, and the design of these (online) places can regularly change towards a higher level of customer satisfaction. One is usually static, while in the other one, there are many possibilities to interact with its environment and learn how to adapt.
Urban morphology meets deep learning: A comparative study of urban forms in 1.1 million cities across the planet. Image @ Vahid Moosavi

What do you think will be the future evolution of data and AI in the design field?

I think the real estate industry and housing will be transformed from a slow and push production system to a personalized, consumer-driven pool system. From the very early stages, the final owners/tenants of a house could get involved in the AI-driven real estate development process and construction and financing. A similar transformation is already happening in the film industry when comparing Netflix’s content production system to the Hollywood film industry.

In the context of generative (structural) design, as explained before, I think AI-based interactive frameworks can lead to personalized design assistants. In those applications that generative tools are not enough or might be restrictive, I believe AI can help architects build their own “personalized catalogs” out of a selected archive of projects. Something like a personalized Pinterest that lives only for a project and is used for a specific case or a specific team.

For architectural offices, I can imagine two scenarios. The first outcome is that large offices will focus only on their internal data sources and their historical records rather than looking into generic and shared datasets or tools. This will create a reinforcing loop, where the large offices can invest in developing highly customized and integrated data-driven frameworks to do faster what they already know well. If this happens, most likely, there will be less information sharing in the industry. The other scenario is that since more designers will have access to more publicly available datasets and cheaper computational capacities, the industry will be more modular with many overlapping technologies that are easy to use.

What do you believe are essential skills to develop for the future generation of architects and designers concerning data and AI?

I think AI is a mixture of a few fundamental research areas that engineering and science students usually learn during their studies. These fundamental areas can be grouped into optimization, linear algebra, machine learning algorithms, statistics, probability theory, and signal processing concepts. Of course, it depends a lot on how to learn these topics, and clearly, and they do not fit the current architectural design programs. However, most existing learning materials are developed with a computer science mindset, and these are not appropriate for architecture and design students.

We should think of AI as a “language”. While computer scientists are more focused on how to develop its underlying mechanics, the architects and designers should become fluent in this language, and beyond “writing” they should think about “what to write?”. This means that designers should be mainly focused on the functionalities of AI and what they can do with it. At the same time, computer scientists as the engineers of these AI systems will focus on the question of how to achieve those functionalities. If we think about AI as a language, it becomes very compositional and with very flexible basic elements (vocabularies) that can be joined together to build a complex application. This is where the designers can contribute better than engineers in joining the basic elements together and articulating complex objects. This way of looking at AI is also aligned very well with the recent developments in deep learning frameworks and the so-called differentiable programming. Therefore, I strongly recommend to architects who want to learn AI to think about it as learning a new language for daily use. Once they can talk in AI terms, the proper design applications will emerge automatically.
Urban Morphology Meets Deep Learning. Video © Vahid Moosavi
Mateo Neira

Mateo Neira

Don't miss data!
Join our Newsletter!

By subscribing you agree with Terms and Conditions, Privacy Policy and all Marketing Permissions. We use Mailchimp as our marketing platform. By clicking on SUBSCRIBE, you acknowledge that your information will be transferred to Mailchimp for processing. Learn more about Mailchimp’s privacy practices here.