Bubble2Floor: An Educational Experience With Deep Learning For Collective Floor Plan Generation
While there has been a large interest in deep learning workflows for generative design over the last few years, most of the existing publications focus on building and reporting research prototypes. In contrast, this paper addresses the use of deep learning in computational workflows for space planning in a pedagogical setting. This is important to increase the number of design professionals around the globe with relevant skills in artificial intelligence (AI), to rethink the curriculum of the profession, and to critically engage in reshaping design technology based on AI with designers in the loop. The paper describes an exploratory research based on a recently created course that introduces different subareas of deep learning to architecture and design students, with a focus on creative and collaborative practices. It presents the description, observations, and results of a design exercise that consists of generating different floor plans of a multi-story and multi-unit row-house complex with an integrated workflow that integrates deep learning, computer vision, and parametric design. Besides being an opportunity for students to learn new skills goal, the exercise raises the following question: what frictions and challenges that emerge when deep learning is used for a collective and parallel design of a single building? The exercise relies on an integrated workflow that comprehends 1) a model based on the pix2pix architecture that translates from bubble diagrams to image diagrams of apartment layouts, 2) a computer vision and clustering workflow to translate from images to geometric entities, and 3) the geometric model of the site, and the building model where the resulting layouts should be integrated to. The groups engaged in the different stages of design with generative models, such as data generation, curation, model training, and fine-tuning, post-processing, and geometric modeling. We report a series of observations on how the students experimented and customized the different stages of the workflow to overcome its limits both to address the specific constraints of the problem (floorplan of a shared building mass) and of their personal preferences. We also analyze the resulting housing layout as a whole and discuss the results. While part of the student work is still under development, the initial results show that the dataset is a crucial factor for the resulting floorplans. It should be flexible or diverse enough to enable students to explore spatial types or crossovers based on personal preferences. Another interesting observation is that a workflow based on image-based diagrams instead of well-structured CAD representation supported the exploration of ambiguous spatial representations and left some space for customization during post-processing. Codes and supporting data will be provided with the paper to support other pedagogical initiatives. References art-programmer. (2017) 2021. Art-Programmer/FloorplanTransformation. Lua. https://github.com/art-programmer/FloorplanTransformation As, Imdat, and Prithwish Basu. 2021. The Routledge Companion to Artificial Intelligence in Architecture. 1st ed. Routledge Companions. New York: Routledge Goodfellow, Ian, Bengio Yoshua, and Courville Aaron. 2016. Introduction to Deep Learning. Cambridge: MIT Press. http://www.deeplearningbook.org. Nauata, Nelson, Sepidehsadat Hosseini, Kai-Hung Chang, Hang Chu, Chin-Yi Cheng, and Yasutaka Furukawa. 2021. “House-GAN++: Generative Adversarial Layout Refinement Networks.” ArXiv:2103.02574 [Cs], March. http://arxiv.org/abs/2103.02574. Nourian, Pirouz, Samaneh Rezvani, and Sevil Sariyildiz. 2013. “A Syntactic Architectural Design Methodology: Integrating Real-Time Space Syntax Analysis in a Configurative Architecturaldesign Process.” In Proceedings of the 9th International Space Syntax Symposium, edited by Y. O. Kim, H. T. Park, and K. W. Seo. Seoul: Sejong University. Weinzapfel, Guy, and Nicholas Negroponte. 1976. “Architecture-by-Yourself: An Experiment with Computer Graphics for House Design.” In SIGGRAPH, 10:74–78.
Keywords: Quality Education; Deep Learning; Collective Design; Generative Design; Space Planning; Industry, Innovation And Infrastructure