3D Generative Adversarial Networks to Autonomously Generate Building Geometry

Abstract

Across the world, countries are facing housing shortages and the Netherlands is no different. The increasing demand for new housing exceeds the growth rate of the architecture, engineering, and construction industry. Current solutions remain small in scale and therefore unsustainable. Multi-family housing is the optimal typology to address the housing shortage but the industry cannot design and build these projects fast enough. Automation can help. In 2016, Kevin Kelly[2016] reframed the conversation about automation, by stating "our most important mechanical inventions are not machines that do what humans do better, but machines that can do things we can't do at all." Humans cannot address the housing crisis alone, but automation through the application of deep learning models can bring well-designed spaces to everyone.

Since the introduction of computers, architects have looked for ways to automate menial tasks. Some researchers even imagine a future where the machine becomes a partner to architects and designers, contributing to the design process. This ambition requires that the algorithms train themselves. Innovations in the field of deep learning have made this possible by allowing algorithms to train themselves through the use of artificial neural networks. When it comes to applying deep learning to generative design tasks, however, there is little research. The studies that have been done generate geometry that is small in size (64 x 64 x 64 voxels) and focuses on objects like chairs, not on buildings. 3D generative adversarial networks show promise for generating building geometry. By automating design, it is possible to apply expert knowledge on good design to all projects so everyone has access to well-designed buildings.

This research aims to develop the architecture for a generative adversarial network that produces feasible building geometry. An important first step was identifying and pre-processing a data set that could be used for this purpose. The data set is released with the publication of this thesis so it can be used for further research. Through this thesis, the Improved 3D Wasserstein Generative Adversarial Network architecture has also been developed and documented. The research found that using a combination of Wasserstein loss with gradient penalty, Leaky ReLU activation functions in the generator and the critic, and the RMS Prop optimizer results in an architecture with stable training and outputs that are similar in size, shape, and proportion to the training data with minimal noise in the output. Performance of 3D Wasserstein generative adversarial networks with these hyperparameters was improved even further when using ten layers and a larger number of channels. The experiments concluded that generative adversarial networks can be used to generate building geometry and can be an area of continued research to improve generative design tools and support the automation of architectural design.