YEAR: 2023
COURSE: M.S. Design Tech (Cornell) Machine Learning Studio
Instructors: Nick Cassab, Panagiotis Michalatos
This project leverages machine learning for the classification and resulting design configuration of architectural elements. Two archetypes of radical and traditional architecture were first selected - Archigram as radical and Palladio as traditional. A dataset of 400 images - 200 Archigram and 200 Palladio - was collected, focusing on drawings and other illustrations instead of photography. These images were used to train a classifier model in identifying these two categories.
Once the model was trained, a series of elements - green toy blocks - were developed to represent each class. These 2 sets of blocks were digitally assembled in countless configurations using Grasshopper with specific constraints. Then, hundreds of those configurations were shown to the classifier model and assigned percentages of "Archigram-ness" and "Palladio-ness." In doing so, some of the configurations made with Archigram blocks began to read as Palladian, and vice versa. Thus, Archigram-coded shapes became Palladian and Palladio-coded shapes become Archigram.
Dataset Images
Image Processing: Images were first converted to black and white and then to outlines to emphasize the texture and drawing style of each. These outline images were used to train the classifier model.
Origins of Block Shapes
Generated Variations (created in Grasshopper): Palladian variations only use the Palladio-coded blocks and are all symmetrical organizations.
Generated Variations (created in Grasshopper): Archigram variations only use the Archigram-coded blocks and are restricted in how they can interact and stack.
The most Palladian configuration with Palladio elements
The most Archigram configuration with Palladio elements - Palladio recoded as Archigram
The most Archigram configuration with Archigram elements
The most Palladio configuration with Archigram elements - Archigram recoded as Palladio
The ML model can "see" in real time which class the given image belongs to.
