The unprecedented development of machine learning (ML) and artificial intelligence (AI) has opened new ways of capturing architectural quality, where large neural networks have demonstrated remarkable capabilities compared to traditional rule-based approaches. However, most AI applications were developed for the general domain, adopted, and repurposed for architectural explorations, thus lacking the capacity to understand the spatial domain in architecture. The growing prominence of architectural visuality further diminishes spatiality in the discourse of AI and architecture.

An apparent missing modality in the current AI approach makes it harder to understand the spatial domain of architecture. This research attempts to fill the gap between the interest in applied AI in architecture and the systematic investigation into the understanding of space. Focusing on the isovist, egocentric spatial perception model commonly used in architecture and environmental psychology, this research investigates representation learning frameworks to capture the spatial patterns of built environments for analytical and generative applications. Therefore, this thesis's overarching question is: What does ML bring to understanding architectural space?

This research investigates the representation learning of spatial domains in architecture by machine learning experiments of isovists sampled from thousands of architectural floor plans. Representation learning seeks to condense essential features of data into a compact, internal representation known as latent space to facilitate understanding and computation. The hypothesis is that the structure and properties of latent space facilitate specific analytical and generative operations in architectural design. These investigations aim to magnify these latent operations, envisioning practical applications and formulating theoretical implications. They illuminate the continuous and discrete modalities of architectural space latent representation and their consequences on the analytical and generative methods for architecture.

The contributions include frameworks for architectural representation learning, shifting the focus from traditional deterministic models to approaches that emphasize implicit and predictive understanding of space. The investigation of latent representations' continuous and discrete operations revealed their affinity to typology, semantics, and decomposition in architecture. At this pivotal moment in AI's adoption within architecture, this research underscores the critical need for domain-specific representation, aligning machine learning with architectural inquiry to advance the theory and practice in machine understanding of architectural space.