Exploring the latent space distribution of a graph autoencoder trained on 3D models of modernist architecture

• Published in: International journal of architectural computing Vol. 23; no. 3; pp. 742 - 752
• Main Authors: Bauscher, Erik, Wortmann, Thomas
• Format: Journal Article
• Language: English
• Published: London, England SAGE Publications 01.09.2025
• Subjects: graph-based architecture; spatial analysis; data augmentation; latent space analysis; graph machine learning; Generative 3D architecture; graph autoencoder; Mies van der Rohe; latent walk
• ISSN: 1478-0771; 2048-3988
• DOI: 10.1177/14780771251353797

Building on previous research in generative graph machine learning in architecture, this paper investigates how data generation and preparation can change the distribution of a model’s latent space and thus its generative qualities. Therefore, we first present and discuss our previous approach of applying generative graph machine learning in architecture by sampling the latent space of a graph autoencoder trained with the augmentations of four examples of modernist buildings. We then present a new method of data generation for modernist buildings in the style of architect Mies van der Rohe, which produces a large range of 3D building models with great geometric variety. Trained on the new dataset, the graph autoencoder shows a more continuous latent space, confirmed by visual comparison and by three spatial analysis algorithms that quantitatively assess the spatial structure of the different latent spaces.