Learning Spatiality - A GAN method for designing architectural models through labelled sections

Abstract

Digital design processes are increasingly being explored through the use of 2D generative adversarial networks (GAN), due to their capability for assembling latent spaces from existing data. These infinite spaces of synthetic data have the potential to enhance architectural design processes by mapping adjacencies across multidimensional properties, giving new impulses for design. The paper outlines a teaching method that applies 2D GANs to explore spatial characteristics with architectural students based on a training data set of 3D models of material-labelled houses. To introduce a common interface between human and neural networks, the method uses vertical slices through the models as the primary medium for communication. The approach is tested in the framework of a design course.