Acqua Alta

Brendan Tsai, Taubman College of Architecture + Urban Planning

Collaborators: Mitch Prasetya Hodiono, Taubman; Fangye Luo, Taubman; Chung-Han (Joanne) Huang, UMSI & Taubman

Artificial Intelligence is pushing how researchers and designers look at the world. By leveraging artificial intelligence’s predictive prowess and optimization techniques, we are looking at reshaping our designed environments. This fresh design perspective allows us to explore various scales, the ontology, the epistemology related to artificial intelligence and the methods to capture these ideas and materialize them.

The appeal of integrating artificial intelligence into worldbuilding can be found in its capacity to materialize alternative architectural narratives and future scenarios that stretch beyond conventional human imagination. The interdisciplinary intersection attracts architects, researchers, and creators as artificial intelligence’s predictive abilities act as a lens into potential future trajectories.

The project Acqua Alta, meaning ‘High Water,’ aims to explore the world of worldbuilding with architecture and artificial intelligence. The project speculates a future of Venice, a city ingrained with historical and cultural significance. Venice, known for its waterways, canals and artistic heritage, faces a threat from climate change.

The project seeks to address this challenge by framing a Venice that has adapted to an aquatic future. By using AI’s capacity for complex data analysis and problem solving, translating these capabilities into an architectural project. This involves using AI to interpret and integrate the city’s architectural and artistic language into new designs that respect Venice’s long heritage while innovatively addressing the challenges of a submerged landscape. The project timeline is between January to the end of April, with an exhibition in May.

Responsibilities are divided in two parts: architecture and technology. The architecture department focuses on modeling, drawing, rendering and representing the project in conjunction with AI. The technology department will explore suitable neural network modes, creating datasets that inform architectural aesthetics. I will facilitate both portions as I have a background in both architecture and artificial intelligence application in architecture, while also ensuring thorough documentation for open-source dissemination.

This project exemplifies the idea of combining multiple disciplines to address global challenges such as climate change. This project, by documenting every step and making methodologies and outcomes open source, will provide valuable learning resources for students as this is a relatively novel approach to architectural design. It fosters a culture of knowledge sharing and innovation in architecture and technology’s role in design.

As of early February, we are largely working with ChatGPT as a novel way to 3D model in the software Rhinoceros. In architectural design, we grapple with what programs fit in what typology, how much area do these programs take up, and how to arrange properly for it to make sense in a circulation. A large language model such as ChatGPT is trained with billions of data from around the world, and we want to leverage this resource. Can ChatGPT provide important information in architecture design, such as coming up with programs for a chosen typology? If so, is it able to distribute these programs?

As of writing the proposal, we found that ChatGPT is able to provide accurate information in regards to the programs. Especially for a complex typology such as Biomedical Research Lab, it was able to produce programs and general flows, or sequence of rooms. We verified these results with existing labs and resources on how to design a Biomedical Research Lab. We then begin to ‘teach’ ChatGPT how Rhinoceros works, how to input script commands, teach it how to structure a command in order to move onto more complex tasks.

A Brief Example (specifics in supplemental information) !_point x1,y1,z1 _point x2,y2,z2 … xn,yn,zn !_point x1,y1 _point x2,y2 … xn,yn !_box x1,y1 z _box x2,y2 z … xn,yn z

These copy-paste scripts allow Rhino to perform a series of commands. These most basic ones allow Rhino to distribute points in the xyz plane and create boxes in the xyz plane.

Another example we designers may be more familiar with would be ! _Select _Pause _SetActiveViewport Top _Rotate 0 30 _SetActiveViewport Right _Shear w0 w0,0,1 -45 _SetActiveViewport Top _Zoom _All _Extents

This is automating the process in Rhino to create Axonometric models. The reason why we want to ‘teach’ ChatGPT how to script in Rhino is because we want to explore the process to automate not only program list generation, but also automate the process of points and box generation to represent these programs in the xyz world.

The goal for ChatGPT is to be able to self determine the ideal size for each program within a given area based on its robust database. Not only the x,y but also the z, or height because not every room requires the same height, different height for different functions. Preliminary results indicate that ChatGPT is able to make these basic decisions and provide lists of programs and sizes in x,y,z coordinates.


Entrance and Reception Area: Z-depth: 12 units (grand entrance) Script: !_box 15,0 35,10 12

Administrative and Support Areas: Z-depth: 7 units (standard office space) Script: !_box 15,10 35,30 7

A portion of the programs given (full version will be in the supplemental list).

It is clear that the result indicates the ability to determine sizes of the program and the height based on necessity. One could imagine an entrance to be more grand and thus require a higher ceiling and a large space. Administrative areas which are more in the backend or support of the building, may not need as high of a ceiling height or overall space.

The next step for the ChatGPT portion is to ask it to put these programs into a predetermined shape or outline to make it more architecturally feasible. To illustrate this, picture a simple, empty rectangular floor outline. We have 4 programs: offices, labs, hallway and bathroom and we want ChatGPT to be able to read an empty rectangular outline and explicitly place these programs in the outline in a way that makes sense.

From my preliminary testing, ChatGPT-4 does have a hard time reading image files (supplemental list will include examples), we are currently exploring more ways to make this possible. This includes but not limited to:

  • Refine image files, perhaps limiting to black and white only
  • Adding a grid system to the sketch, because we know that ChatGPT can work with grid, but this method with geometry becomes far more complex
  • Text description

Graphics are an important part of the project, our goal is to scrape various images of Venice or broader Italy for representation, both three dimensionally and two dimensionally. As of now, we have not done much with this process except finding a way to scrape images.

Using code executors such as Google Colab, we can find base codes for tasks such as: Image Scraping, StyleGAN2, StyleGAN3 and Latent Walk…etc. The obstacle at the moment is finding a script that works properly, StyleGAN is incredibly old, and some of the host files that we tried are either removed or outdated. We hope to collaborate with computer science students to modify these codes to suit our needs and make it work.

By feeding Neural Networks thousands of images of various topics, take facades of Italy as an example. The Neural Network is able to interpret new facades of Venice that we can use for the project. There are methods to translate 2D image to 3D texture or models that we can work with in conjunction with the earlier process described above, which will result in a more robust project.

The ChatGPT portion has a much more robust output because it is something we have been working on for a while. This portion of the project hasn’t been touched much yet because it’s currently too expensive to work with.

As part of the project, we are expected to have 2 exhibitions. One in late April and one in early May, at Taubman College and the Liberty Research Annexe. An important part to advance AI in architecture is to be able to build your project. We are expecting to build at least one large model, perhaps several smaller chunk models to represent our AI work.

To complement this, we are also expecting to print a series of large scale drawings, renders and other representations to illustrate the project. Most importantly, all our research, methodologies, code, and sources will be made available open-source in a book format.