AI Generative Design
Why Do Building Designers Need AI?

Why Do Building Designers Need AI?

From chat-bots to alpha go, AI is becoming increasingly personalized and accessible in previously unthinkable ways. But why do building designers need artificial intelligence?

From chat-bots to AlphaGo, Artificial Intelligence (AI) becomes increasingly personalised and accessible in previously unthinkable ways. Today, even if you aren’t a computer scientist, with some training, persistence, and a good WIFI connection, anyone can create their own AI system to help them with their job.

As an architect, here are three ways I have been using AI to process information, become more creative, and develop better building designs. 

IMAGE 01 - Illustration of Automated Design Model. Source: Author’s Documentation

1. Learning Complex Pattern and Principle

Designing is not a linear process. It involves humans as its user with social, economic, historical, and many more aspects that eventually stimulate the results; hence, design, as a result, is never the same in every context. So how can we automate the design process if each is unique? Here is where Machine Learning can help us with the design process. Machine Learning can learn the principle of specific tasks by feeding it a lot of data. The particular case can learn the particular input pattern and process it as a database, not just only the images. Hence, it is a great strategy to utilize it as part of the design process.

One of the samples currently developing in our research and development team is CFD prediction. CFD (Computational Fluid Dynamic) is a field of computation that simulates wind movement in buildings or any design. In our research, we used the Pix2Pix algorithm to do the training. Pix2Pix itself is an algorithmic model that does image translation, meaning that when we input an image we can have the result of another embodiment. Pix2Pix is also called supervision learning because it needs a pairing of an image, as input, and the wanted output. The algorithm will learn some principles and how the input relates to the wanted output. In our samples, the training is a pair of site samples with building mass inside it, and the expected output image is the wind analysis pattern on the site. We expect the machine to learn how building mass within the site can develop or affect the wind flow pattern to predict where high wind pressure and speed will occur.

IMAGE 02 - Illustration of Machine Learning of Heatmap Process. Source: Author’s Documentation

After the training, we can generate many iterations for building typologies and, therefore, be useful for designers. This proves that Pix2Pix, or Machine Learning in general, can learn something complex, with calculation inside it, geometry, even context, wind direction, and wind speed in pairs of images. The results can be used to create another design proposal in less time (after training).

Another exciting thing about Machine Learning applications is they can transfer a particular style of an image into another style that does not exist yet. One example is the image of the organic city plan below.

2. Bring New Possibilities of Creative Process

This transformation aims to transform the existing grid pattern or city pattern into another pattern; in this experiment, it transforms into an organic shape based on a natural order pattern. How do we do that? First, we collected images of natural patterns in our world, mountains, deserts, even forests. The machine then trains this data to learn the pattern by reducing the data into black and white images to learn the differentiation of the pattern. By having this kind of data, the machine knows how to structure the pattern based on a monochromatic image. If we input a new image here, the machine will let the input read as a black and white image and replace this transformed input with the black and white pattern in training Thus the input will transform its existing image structure into a new output structure based on the training.


IMAGE 03 - Illustration of Machine Learning that can transfer the style of a certain pattern and change the city pattern model. Source: Author’s Documentation

In this case, Machine Learning can help us rethink what we did not think before. As in this experiment, We can imagine how to transfer the city's existing grid into a new organic pattern of the city. By understanding this, we can prove that machine learning can help us to rethink our creative process. It broadens our perspective and imagination while giving us another possibility of a solution that maybe is usually hard to imagine.

3. Faster and Easy Prediction

We already know that Machine Learning training takes quite a lot of time and effort. In some cases, we need a lot of training data, not only hundreds but thousands - with hours, days, and even weeks' worth of time. We also need to filter and select which data is suitable for training, and we need a lot more time to label each data to create coherent input-output training pairs like CFD Machine Learning above. Then after training is finished, it is when we pick the result.

 IMAGE 04 - CFD Pattern Prediction using Pix2Pix. This video was a time simulation using our development hand sketch apps. We can see that every footprint given, it can directly generate the CFD pattern. Source: Author’s Documentation

For example, from previous sections, the CFD, the data collection and generation takes time. Still, after we have the training model, we can analyze many contexts in many projects in just seconds. We can do it in many sites, many iterations of forms, and it will be only in seconds. The analysis will be swift, and we can allocate the remaining time to iterate forms, programs, and other essential parts of the design. In this aspect, ML doesn’t give us the direct benefit of having creative results. However, It saves us time to help us think critically about our design, re-questioning our entire process with the help of Machine Learning, and thus indirectly enhancing our time to implement more creative solutions.

What’s next?

IMAGE 05 -Illustration of Generative Design combined with Heat Map prediction using Machine Learning. Source: Author’s Documentation

AI cannot be a solution for everything; it is not bulletproof. But indeed, nowadays, AI is still an open field every day, developing and growing. DBF currently focuses on generating automated designs for a sustainable built environment, it can help enhance the creative process and create the possibility to give solutions for the company's vision.

We know that it cannot guarantee 100% good results, and AI still gives us a lot of homework to do, such as how we can save time yet still be accurate in collecting and producing the database for training, how to evaluate the result and optimize it for a more usable result, or how to combine numerical data and image data for optimizing the training model, and so on. This homework, we believe, is not only for us, but for all of the architects, urban designers, and anyone who is pushing the boundaries between AI and Architecture; it is our job to do it together.

How do we know what AI can do for us in the future? We think AI can be more precise yet meaningful for our design process. AI can help predict structural patterns that reduce material waste or predict material placement in a building that optimizes human comfort. Currently, it is mainly about the creative process, but in the future, it will be about helping create a sustainable environment in the future. We will see.


  1. Chaillou, S. (2019a) AI + Architecture | Towards a New Approach. Harvard University. Source: https://www.academia.edu/39599650/AI_Architecture_Towards_a_New_ Approach. 
  2. Chaillou, S. (2019b) ‘The Advent of Architectural AI - A Historical Perspective’, Towards Data Science. Source: https://www.academia.edu/39802404/The_Advent_of_ Architectural_AI_A_Historical_Perspective
  3. Joler, V. and Pasquinelli, M. (2020) ‘The Nooscope Manifested: AI as Instrument of Knowledge Extractivism’, Nooscope.ai. KIM HfG Karlsruhe and Share Lab, pp. 1–23. Source: https://nooscope.ai.
  4. Isola, P. et al. (2016) ‘Image-to-Image Translation with Conditional Adversarial Networks’, CoRR, abs/1611.0. Source: http://arxiv.org/abs/1611.07004
  5. Pasquinelli, M. (2019) ‘Three Thousand Years of Algorithmic Rituals: The Emergence of AI from the Computation of Space’, Journal #101 - e-flux. Source: https://www.e-flux
  6. Steenson, M. W. (2017) Architectural intelligence : how designers, tinkerers, and architects created the digital landscape. Cambridge, MA: The MIT Press ; London

Join our waiting list

Be the first to get exclusive access to early releases, valuable insights, and platform updates from DBF.

By submitting this form you agree to receive emails, text and promotional communications from DBF related to products and services in accordance with DBF privacy policy.

Thank you! You have joined the waiting list!
Oops! Something went wrong while submitting the form.