Augmented intelligence for architectural design with conditional autoencoders: Semiramis case study

We present a design approach that uses machine learning to enhance architect's design experience. Nowadays, architects and engineers use software for parametric design to generate, simulate, and evaluate multiple design instances. In this paper, we propose a conditional autoencoder that reverses the parametric modelling process and instead allows architects to define the desired properties in their designs and obtain multiple predictions of designs that fulfil them. The results found by the encoder can oftentimes go beyond what the user expected and thus augment human's understanding of the design task and stimulate design exploration. Our tool also allows the architect to under-define the desired properties to give additional flexibility to finding interesting solutions. We specifically illustrate this tool for architectural design of a multi-storey structure that has been built in 2022 in Zug, Switzerland.