Josef Musil
Josif Musil, Associate Partner – Design Systems Analyst, Foster + Partners | Image © Foster + Partners

AI in Design: A Conversation With Josef Musil on the Use of Machine Learning in Architecture Design

Josef Musil is an Associate Partner at Foster + Partners in London, where he is part of the research and parametric design-oriented Specialist Modelling Group. His work focuses on applied research, application of new technologies, and algorithmic design to complex architectural and geometrical challenges often in connection with environmental design and analysis. Josef studied as a Fulbright scholar at the University of Pennsylvania, School of Architecture, where he received a research MArch degree on the topic of computational design, and is enthusiastic about bridging computer science and other sciences with architecture. He has also worked as a researcher and tutor at UPenn, US, UCL, and AA visiting school.

During his time in Foster + Partners he has worked on projects ranging from the facade screen for the Edmond & Lily Safra Center for Brain Sciences at the University of Jerusalem, where he built a generative model for small section of the brain along with an automated workflow that reads data form 3D neuroscientific scans and implemented structural analysis results and manufacturing constraints, to developing a working functional prototype of a small number of 3D printing swarm-based robots for a NASA organized Mars habitat competition. Currently he is looking at implementations of machine learning algorithms into architectural design process and simulation.

One area that is prime for the use of machine learning algorithms in design is in the development of surrogate models that can be used in early design stages to quickly iterate through different design options and get quick feedback about its performance. A surrogate model is a simpler model of a slower computational process (such as a physics-based simulation) and is usually developed to learn nonlinear relationships between input and output values of the original model. Musil has been working on developing these sorts of surrogate models to enable interactive fast design cycles and reverse typical design workflows. We spoke with Musil about how he uses Machine Learning as part of the design process and how surrogate models can help better understand different design constraints.
Image © Foster + Partners

How do you implement data and AI in the design process in the projects you work on?

We try to use as much data as possible to drive any design decisions. The most common data comes from environmental analysis in initial design stages and material properties in later stages. We used AI algorithms to generate high resolution material textures, which is very useful, if the material is the main aspect defining the look of a large area. Natural materials do not have any repetitions but have small characteristic details and a generative adversarial network helped us to achieve both. We also used a convolutional neural network to predict CFD (Computational Fluid Dynamics) analysis during initial stages, orders of magnitude faster compared to a full analysis.

What are the opportunities in using data and AI in your design process?

I am very much interested in using AI algorithms to reverse typical design workflows, like the reverse CFD workflow, presented here. We can generate lots of data in the forward direction, e.g. generate lots of random design options and compute full CFD analysis for all of them. Using this data then to drive the design is more difficult. Also, such an analysis is often seen to evaluate the design, but air flow has a direct impact on how people perceive spaces and architecture. We thus want to design those effects like wind flow directly, rather than only check what effects a certain design produces. As there is currently no direct way to go from wind analysis to corresponding architectural geometry, AI algorithms will provide a unique opportunity to do so. The same can apply to lighting or sound for example.
Image © Foster + Partners

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

In this specific case of predicting CFD using 3D convolutional neural networks, the main challenge is to minimize the amount of simulations needed to create an effective training dataset. That means finding a way to make the most effective sample selection out of the whole continuous range of designs, when using synthetic data generated by a parametric model with a high number of parameters. This is to reduce computation time to prepare the dataset but also more importantly to create generalization good enough across the whole range of possible designs. As architecture is often based on creating a unique design, generalization across projects is even more difficult. Transfer learning should provide help with this. Regarding the reverse direction to predict architectural volumes based on desired wind flow, to effectively create the training set is even harder as we can only start with samples of the architectural volumes not the wind analysis.
The speed of the prediction not only allows us to evaluate large number of options but also facilitates the communication of the results through animations which eventually helps to understand the design constraints better.

Where do you see the future evolution of the use of data and AI in the design field heading?

Hopefully we will soon see more generative algorithms working in the field of graph based networks which allow working with irregular spatial data. That will allow us to generate meshes directly rather than voxels as we show in our CFD example. Or to generate urban plans and urban relationships or internal room connectivity, which is all based on irregular patterns. Generative models will be hopefully advanced by the ability to be directly manipulated in feature spaces.
Image © Foster + Partners

Could you talk about one of your latest projects that make use of innovative design workflows and how AI was used ?

We have used our AI CFD tool a few times at concept design stage, which proved to be very useful as we could study the impact of major mass changes of air flow and keep up with the fast development of the design typical for early stage of a project. The speed of the prediction not only allows us to evaluate large number of options but also facilitates the communication of the results through animations which eventually helps to understand the design constraints better. Our interactive CFD application, which can also be seen as a game that helps its user to learn about the relationship between geometry and wind flow, helps to create an intuition about future design changes and their impact on wind flow.
A convolutional neural network to predict CFD (Computational Fluid Dynamics) analysis during initial stages of design, orders of magnitude faster compared to a full analysis. © Foster + Partners
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.