Designing Small Datasets: ML modeling applications for 3D printed biopolymer composites in architecture

This thesis investigates methods for the composition of small datasets aimed at leveraging ML algorithms for implicit modeling of biopolymer 3D printing for architecture. The research is contextualized against the need for an urgent shift in architecture’s material practice facing the challenges of sustainability, and the inability of the current digital modeling paradigm to sup-port this transition. This is because bio-based materials are fundamentally different to their industrial, standardized counterparts presenting complex, heterogeneous behaviors. There-fore, this thesis proposes to extend contemporary understandings of the Complex Model by integrating implicit ML-models within the larger non-predictive computational pipeline. This integration provides an alternative method for complex material system modeling. It leverages the potential for rapid and focused predictions that can be seamlessly incorporated into the design environment, thereby augmenting the model with functionalities that are unattainable through alternative modeling techniques.
The research operates at the intersection of the architectural field of computational design and fabrication, and the mathematical field of scientific modeling using ML algorithms. ML-workflow building is framed through a Research through Design lens, and is the central method taken up in this thesis. It is applied across experimental case-study work, which is characterized by iterative, feedback-based, multi-staged Test-Driven Development. The aim of the research is to understand the relationship between small datasets and ML models.
The novelty brought on by this thesis is the focus on data, dataset creation, and data flow inte-gration, which are essential to the deployment of ML in our field. It points towards a change in tooling culture, where the focus becomes on the transfer and adaptation of existing algorithms, fed with specific data, rather than the development of the algorithms themselves. The research focuses on issues of data scarcity and implicit complexity, inherent to the nature of design and architecture. It demonstrates methods for small dataset curation and ML-workflow design which mitigate such application challenges.
The thesis is structured across two digital modeling experiments which test two opposing methods of designing small datasets. The two experiments are anchored within two case stud-ies, which frame the material case and the functional motivation for the modeling experiment's target. The case studies address the critical modeling limitations which hinder the usage of bio-based material, namely: modeling the dynamic curing of the 3D printed biopolymer com-posites, as well as the characterization of graded 3D printed biopolymer composites. A large ensemble of methods is utilized across the case studies: methods of digital parametric design, digital fabrication and robotic control, 3D scanning, digital sensing, data structures and han-dling, data processing, graphical and numerical data analysis, model training and finally meth-ods of visualization and user interaction.
The findings of the experimental work contribute to knowledge within three territories. 1) The thesis advances the current limits of architectural digital practice in modeling biomaterials, by producing a novel method for complex material modeling integrated within the computational design model. The method enables the navigation of a recipe space, through the simultaneous coupling of material composition and material properties. 2) The thesis extends the current understandings of architectural Machine Learning in terms of datasets and algorithms by de-veloping methods for custom dataset creation. The method enables to sample complex design spaces, and through numerical encoding, transform design-driven physical demonstrators into machine learnable datasets. 3) The thesis expands the methodology of Research through De-sign (RtD), and the framework of Probe-Prototype-Demonstrator to the act of ML-workflow building itself.