Ai + Architecture Thesis
University of Michigan | Taubman College of Architecture
Spring 2020
Collaborator: Peter Recht
Architecture and Artificial Intelligence presents itself as an opportunity to critically investigate the role of AI in a future world of building, including the conversations on the impact as a cultural technique for the production of architecture design. The objective of this thesis is to train a Neural Network on 3D building features to generate novel point cloud outcomes using Unsupervised Machine Learning. This thesis interrogates the consequences of the implementation of AI techniques pertaining to its role as an agent of culture. 
Data Base Generation
Original Building Database This thesis began by analyzing the site’s climate data. Buildings with similar climatic parameters were added to the database. A database consisting of over 1,000 models has been modeled.
Database Generation The original building (left silhouette) were morphed using Rhino Cage Edit (right silhouette) to create iterations for training. The iterations between the models were blended to increase the size of the database.
Unsupervised Machine Learning: Tree-GAN 3D Point Cloud Generative Adversarial Network Based on Tree Structured Graph Convolutions
Pipeline of the Tree-GAN The tree-GAN contains two networks, namely, discriminator and generator. The generator takes a single point from a Gaussian distribution, z 2 R96, as an input. At each layer of the generator, GraphConv and Branching operations are performed to generate the l-th set of points, pl. All points generated by previous layers are stored and appended to the tree of the current layer. The tree begins from the root node z, splits into child nodes via Branching operations, and modifies nodes by GraphConv operations. The generator produces 3D point clouds x0 = pL 2 R3n as outputs, where pL is the set of points at the final layer L and n is the total number of points. The discriminator differentiates between real and generated point clouds to force the generator to produce more realistic points. We use a discriminator similar to that in the r-GAN. Please refer to supplementary materials for detailed network architectures.

Loop Term with K-supports: (Left) A conventional loop term uses a single parameter to learn mapping; (Right) The loop term introduces a fully connected layer with K nodes to learn a more complex mapping

Example of Branching with Degree 2

Ancestor Term A conventional neighbor term uses neighbors to generate

Ancestor Term The Posed ancestor term uses ancestors to generate

Preliminary Results- In Progress

Example of the Point Cloud Data Generated using the Neural Network

Systematic Study of Results Testing results from 300th epoch (about 30% of overall training) all the way to 100% training. We are trying to see if patterns emerge within point clouds at different stages of training (horizontal rows) or within point clouds across different stages of training (vertical columns)

Development Across Epoch
Development Across Point Cloud
Representation - Cubes - Point Cloud 3
Representation - Cubes - Epoch 600
Back to Top