Rematerializing Graphs: Learning Spatial Configuration

In architecture, recent developments in machine learning technologies have reenergized the established research areas of generative design and design analysis. Proponents argue that data-driven systems can learn the implicit rules of architectural design for the generation of new designs or the incorporation into new methods of architectural research and analysis. This recent research has focused primarily on machine learning model development, in particular the algorithms, but not the data from which the models learn. However, without a concurrent reflection on the data and the representations used to encode the data, a critical understanding of the affordances and limitations of data-driven machine learning tools in architecture is not possible. Therefore, in this research I interrogate the data with a focus on graph-based representations of spatial organizations. As a research method, I present the theoretical concepts tying graphs to computational design theory to selected cutting-edge graph-based and data-driven generative and analytical approaches. I also reconstruct an example of early generative design software and create computational instruments to explore, demonstrate and evaluate the concepts presented in the literature research. Placing recent analytical and generative techniques in the lineage of graph-based representations demonstrates how the motivations of the early theorists, and limitations of graph representations themselves, are carried forward into the machine learning context. Further, investigating the floor plan data sets essential to these technologies reveals the inherently contingent nature of graph-based architectural data, and by extension, the contingency of the results.