Since its conceptualization in the early 60s, the role of computational tools in creative inquiries has been a hot topic of debate. In contrast to the early visions to create collaborative partners, contemporary computer-aided design (CAD) systems heavily favor the computer’s role as a task executor. However, creative inquiries are rarely task-oriented and often involve reciprocity between seeing and acting on emergent properties.
Within this problem definition, the overarching goal of this research is to investigate learning-based models trained on problem-specific datasets to facilitate creative inquiries by deriving design suggestions from otherwise hidden latent space based on partial or incomplete design solutions. This paper presents the first stage of this research investigating the ability of Autoencoder (AE) networks to represent and learn different design approaches and intentions in lower-dimensional latent space and its reconstruction capabilities through problem-specific synthetic datasets.
Collaborators: Shakthi Suresh, Chris McComb