DEEP WINNING FORM: Machine investigation of architectural quality

This paper showcases the development of Arch-Form, a plat form that enables the investigation of underutilization of knowledge from architectural competitions, specifically within the Swiss architecture system. The aim is to leverage machine learning to analyse and understand architectural forms from school competition data spanning the past 20 years. The original contribution of this study lies in transforming competition results into a machine-learnable format, using 622 massing models to create 'architectural' point clouds. This methodology involves using 3D Adversarial Auto encoders (3dAAE) to encode and reconstruct these point clouds, experimenting with various structured formats such as uniform, horizontal and vertical g-codes. The main conclusion drawn is that machine learning can significantly aid in understanding and predicting architectural form preferences, documenting trends, and transformations in design. This approach enhances the computability of architectural forms. It offers a new perspective on how machines interpret and generate architectural data, contributing to a more comprehensive understanding of architectural evolution and societal preferences in design.