Latent Design Spaces: Interconnected Deep Learning Models for Expanding the Architectural Search Space

This work proposes an adoption of Artificial Intelligence (AI)-assisted workflow for architectural design, to enable the interrogation of possibilities which may otherwise remain latent. The proposed design methodology hinges on the “systems theory” consideration of architectural design, expressed in Christopher Alexander’s “systems generating systems”, offering an alternative to the reductionist and complexity-lacking structure of design processes (Alexander 1968). This logic is pursued through the integration of DL models into an “open-ended” workflow of interconnected Deep Learning strategies (DL) and other computational tools, rather than treating it as a closed “input–output” cycle (single DL model). While a closed cycle risks flattening architectural layers, ending up with a reductionist encoding of design intentions, an open-ended workflow can inquire into an expanded design search space and augment creative decision making. This way, chained model strategies can simultaneously address design intentionality within discrete architectural layers (i.e. organization, composition, structure).This system enables three distinct modes of collaboration: human–human, human–AI, and AI–AI. Understanding the contribution of human and machine agents within the workflow offers a re-evaluation of designers’ processes. The proposed nested workflow reflects the transition from ‘expert systems’, which rely on hard-coded rules, to ‘learning systems’, which are inspired by the human brain (DL) (Hassabis 2018). This allows architects to approach design problems which are not fully defined (Rossi 2019) and helps avoid over-constraining the search in creative domains like architecture. Furthermore, the design investigation is strengthened by accessing a search space that is otherwise beyond the designer’s reach towards an expanded design creativity.