Project Summary: 

Early-stage architectural design increasingly uses computational tools while structural reasoning remains a post-hoc evaluation. Although ML has increasingly been applied to structural performance prediction, optimisation, and generative design, its use as an interactive form-finding mechanism within early-stage architectural design remains comparatively limited. This research investigates how ML can support structural form-finding as an active part of early-stage architectural design. 

The research has developed through a series of investigations. A literature review and industry engagement with COX Architecture highlighted the impact of structural material on embodied carbon and the potential of computational workflows to predict performance. Two exploratory hackathons further tested this issue. The first investigation explored whether ML could learn structural and embodied-carbon relationships from solver-generated data. While this demonstrated potential for performance prediction, it also revealed that learning simulation outputs does not necessarily mean learning structural logic. The second introduced graph-based representations to support relational reasoning, but revealed that connectivity alone does not encode equilibrium. These findings established the need to represent structural behaviour explicitly rather than only through analysis results. 

This direction was then tested and exhibited through prototyping a project at the Digital Futures international conference, where I designed structural models using Polyhedral Graphic Statics (PGS). The process demonstrated the potential of graphical form-finding methods, but also highlighted the materiality gap, as the method does not directly determine material selection or member thickness, requiring additional analysis to translate forms into materially resolved structures. 

These findings inform the next phase of the research, which investigates a computational framework that extends PGS from geometric equilibrium toward adaptive structural materialisation. It integrates finite element analysis, volumetric modelling, Signed Distance Functions (SDF), and physics-informed ML to relate geometry, equilibrium, and structural response to spatially differentiated thickness adaptation. Rather than seeking a universally optimal material distribution, the framework is conceived as a decision-support system that generates condition-dependent material responses while preserving opportunities for designer judgement and reinterpretation.