Poké-GAN

Date: 2019, Spring

In collaboration with: Qin Xiong, Anthony Samaha, Lin Liu , YiQun Wang

Advisors: Casey Rehm

The project utilizes convolutional neural networks in combination with robotic assembly algorithms to produce scaled timber structures.  In addition to 3D form generation.  Neural networks will be used in to develop 2D texture applications for the structures, exploration the relationship between 3D form and 2D image.  The project explores the opportunities and limitations inherent in machine learning based automated design and fabrication.  Additionally, the project explores how mereological explorations within architectural production can be utilized as generative inputs into non-human forms of graphic expression.

 

The project has three main components.  The first part focuses on training 3D GAN convolutional neural networks on student created datasets of Pokemon files to produce massing models of proto-architectural objects.  These models are then translated into pattern schedules for wood tectonic elements and assembled utilizing the robots.  Pattern and tool path generation utilizes Grasshopper and Python with initial techniques utilizing cellular growth algorithms.  The project subdivides massings into rational sub-components to accentuate pattern in relationship to form and optimizes orientation for the robot work sphere.

64 3D RaSGAN (Geometry)

CycleGAN (texture)