Unsupervised classification of images in generative computational design processes

Abstract

The use of computational techniques in Generative Design processes allows for the creation of numerous solutions. To explore these iterations, it is necessary to adopt approaches that help organize and analyze this vast solution space. This study proposes a deep learning-based methodology for unsupervised image classification, enabling the automatic exploration of generative system iterations without the need for manual labeling. The research uses the Grasshopper software to generate the iterations and transform them into images, a convolutional autoencoder to extract compact representations of the images, dimensionality reduction techniques such as UMAP, and the Gaussian Mixture Model algorithm for data clustering. The work demonstrates that this approach can organize the solution space based on morphological and spatial characteristics, contributing to new forms of interaction between the designer and artificial intelligence in the design field. The codes and images used are available here: https://github.com/researchai2/cluster2.