In the context of ongoing research on incorporating deep learning (DL) strategies in architecture, this paper proposes a proof of concept, for developing a viable DL-driven design workflow with multiple connected DL models that enable various levels of agency. The approach allows design intentions to manifest systematically throughout the process, through identifying the ways of dataset curation, DL models' selection and connection. Importantly, in parallel to the interconnected DL models, a series of physical experiments were conducted for dataset augmentation and evaluation, and to inform the overall process. The formulated system involved protocols where multiple DL models are employed and interconnected to address specific architectural systems and design tasks. Applying this prototype, a test-case experiment was carried out with a parallel logic of the two processes: (1) a physical experiment (material research) and (2) the DL-driven process (a combination of multiple neural networks), incorporated into the design workflow. The physical experiment was directed at learning from fungal natural systems (mycelium) to understand growth behavior and its physical qualities, which influenced the DL testing and evaluation. 

Read the whole paper: [y/n]?