ITERA N-1(17)

Alejandro Romero – Amisha Bavadiya – Jahan Selim – Selin Dastan

In fractal generation, each step in the process of transforming a starting point or shape into a more complex pattern is called an iteration, represented by “n”.

Study Area for the City of Yes Midtown South

Study Area for the ‘City of Yes’ Midtown South Mixed-Use Plan

The Midtown South Mixed-Use Plan aims to create housing opportunities and increase demand for commercial office space, fostering city growth.

This is our project site, and we chose it for several reasons. Firstly, it permits 2,300 homes through various zoning rules. Furthermore, its central location, excellent transit access, and proximity to strong job centers make it an ideal location for housing development.

Another reason is the potential to add 5 more floors, based on FAR (Floor Area Ratio) map information, which influenced our decision to choose this area.

The building selected for the project. Currently, part of the building is used for commercial purposes and part is empty. It is a perfect choice for conversion to a new housing project.

The design of the Tomb showcases his profound engagement with geometric forms, the adept manipulation of light and shadow, meticulous articulation of architectural elements, and a rich exploration of textures and contrasts.

He artfully combines different materials such as wood, stone, and uses their unique textures and properties to create spaces that are both visually striking and harmoniously integrated with the surroundings.

After exploring Carlo Scarpa’s work, the three images formed the core concepts for our project. The first image gave us an idea to play with mass and voids. The second image, which looks like a crystal, highlights Scarpa’s detailed and repetitive geometric style.

And by looking at the pavement designed by Carlo Scarpa at the Museum of Palazzo Querini Stampalia in the third image, we were inspired by the simple yet effective rules he used. Scarpa’s approach reflects a computational mindset, showing an indirect connection to cellular automata (CA). While we don’t intend to directly copy Scarpa’s work, we thought about how a rule-based system (CA) could guide machine learning to create design.

Cellular Automata

Cellular automata are like a grid where each cell, changes based on simple rules from the cells next to it. This grid can keep changing forever, creating different patterns of mass and voids.

Mass and Voids

The initial plan demonstrated how architects create mass and voids, prompting us to apply cellular automata (CA) to enhance this process. Initially, we generated a CA configuration (random state 0) which resulted in a pattern with more mass and less void. We aimed for a configuration with less mass and more voids. Subsequently, we modified the CA with the existing walls, achieving the desired outcome. The last image depicts how machine learning predicted the distribution of mass and voids.

Pix2Pix Algorithm Model

In the Pix2Pix model, both the input image and the ground truth are used to train the machine learning system. The model learns from this dataset, utilizing a generator to create predicted images and a discriminator to evaluate how realistic they are.

We began by creating our own datasets of architectural plans. We focused on selecting plans with the service core in the center, matching the layout of our existing building.

Data sets – Original Plan

This dataset consists of the outer boundaries of the architectural plan, depicted as completely black areas. These boundaries indicates the maximum or minimum extents within which architectural spaces can be placed. This dataset helps the ML model understand the spatial constraints and the permissible building area.

Data sets – Combined Section

We superimposed the Cellular Automata (CA) representations of mass and voids onto architectural sections, crafting a distinctive dataset. This dataset will be used to train a machine learning model, allowing us to analyze the outputs it generates and assess its spatial planning capabilities.

Predicted Image Generated by Machine Learning -Plan

The first image is the original boundary that we supplied the machine learning system, and the other is the ground truth, which is the merged architectural plan with the CA. The output is the predicted image where we can see how machine learning distributed the mass and voids and the architectural spaces.

Predicted Image Generated by Machine Learning -Longitudinal Section 

Above are several examples of 3D models which features materials such as concrete, wood . These models draw inspiration from Carlo Scarpa’s distinctive approach to combining materials in architecture .

The question was how to inhabit the spaces, as it was clearly visible from the 3D model that there was a problem with mass—it had more mass and fewer voids.

Data sets – CA for Plan

So, we tried again to generate the Cellular Automata based on the existing walls, aiming for less mass and more voids. The result was exactly as we desired.

Data sets – Combined with Existed Plan and CA

Data sets – Plan

This is the combined architectural plan and the new Cellular Automata model with less mass and more voids. We fed it into the machine learning system to see how it would generate the output.

Data sets – Combined with Existed Section and CA

Data sets – CA for Section

Predicted Image Generated by Machine Learning for Plan

The first image shows the original architectural plan, the second displays the mass and voids generated by Cellular Automata, and the third is the image predicted by the machine learning model.

Predicted Image Generated by Machine Learning for Section

Application of the project