Exploring Architecture in Morphology through Machine Learning

Sammanfattning

The application of machine learning in architecture has raised important concerns regarding creativity, authorship, and the usability of AI-generated outputs. An existing limitation is that many generative workflows produce visually compelling images but they remain as visual residue. They may demonstrate rich architectural qualities, but they are detached from the design workflow. This thesis explores how machine learning can be used as a form-finding medium to translate abstract morphological patterns into architectural form and editable three-dimensional design artefacts, while retaining the designer’s authorship throughout the process. The research begins by taking inference from architecture design process, which often starts with an abstract sketch and gradually evolves into conceptual form. This thesis follows a research-through-design methodology structured across three sequential phases: Creation, Transformation, and Generation. In Creation phase, synthetic morphological patterns are generated and hybridized to mimic the unpredictable and complex morphologies of nature. In Transformation Phase, the hybrid morphological patterns are translated to architectural forms using a machine learning model for image to-image translation. Instead, of training the model to directly imitate the existing buildings, the transformation is learned through a custom dataset curated around architectural formal qualities. In Generation phase, selected two-dimensional outputs are generated as three-dimensional model by integrating 3D reconstruction model directly inside architecture design tools. This research demonstrates that the abstract morphological patterns with no direct architectural origin can be successfully translated into architectural forms and finally to editable three-dimensional geometry. The generated forms carry architectural qualities such as surface articulation, tectonic character and spatial depth while still preserving the original morphological identity. However, the outputs remain conceptual form-finding artefacts rather than complete architectural proposals. This thesis proposes a grey-box workflow: a structured planning of existing machine learning tools in design process which places the designer as curator, evaluator, and decision maker at every stage. It positions machine learning not as an autonomous system, but as an expert collaborator for exploring architectural formal possibilities.