The thesis is carried out within the framework of the GenH2Arch research project, funded by the AURA Region (2021-2026) and the AIA Life Designers agency. It takes place both in research laboratories and in the agency. It is supervised by Serge Miguet (LIRIS laboratory, IMAGINE team) and Xavier Marsault (MAP-Aria laboratory). More details on the GenH2Arch project can be found here.

The thesis will mainly take place in the MAP-Aria laboratory, École Nationale Supérieure d'Architecture de Lyon, Campus de Vaulx-en-Velin (main employer), but also partly at LIRIS and at the AIA agency in Lyon.

Details :

The health-friendly urban planning (HFUP) movement aims to promote a holistic consideration of health issues (physical, mental, social well-being) from the earliest design phases of development projects [CAP 18]. The thesis questions how artificial intelligence can support the design of architectural projects in these phases, guiding designers towards the smartest possible solutions from the start. It is based on the alliance of deep learning and generative design for an architectural and urban design favorable to health (specific proposal of the GenH2Arch project).

A great interest of deep learning is indeed in the initialization phases of the design of an architectural project, when the available data are insufficient both to produce morphologies and to evaluate them. Its predictive and generative abilities can then be called upon to overcome this lack of direct knowledge, once complexity has been captured and learned from many more or less similar projects or artificial data produced on purpose. This is the structuralist approach of the project: a structural and functional description by graphs [AS 18] is of particular interest to us in the context of this thesis, calling on geometric deep learning [BRO21], a recent field of study to generalize the techniques from deep learning to non-Euclidean data frequently used by architects.

Decision-making in terms of architectural morphogenesis (forms and interiors) in a given site can then be greatly facilitated by processing this knowledge accumulated upstream, and which generative design allows to be optimized in multiple ways (referring to EcoGen, software developed in recent years at MAP-Aria [MAR 19], with the intention of coupling it to AI tools).

At the local level of a block or a group of buildings, the health approach will be broken down into a range of specific determinants and indicators that will feed the hybrid generation. The case study of hospitals is of particular interest to us; it will draw on AIA's expertise in this area and on evidence-based design.

The fundamental importance of the relationship to the site (no good off-site analyses) questions the association between buildings and context (deep learning methods based on graphs and image data). At this exploratory level is raised the question of trying to obtain a good conditional generation of sketches or project scenarios in a given site (natural and built environment), responding to a program and structural and functional constraints.

References

[AS 18] I. As, S. Pal, and P. Basu, “Artificial intelligence in architecture: generating conceptual design via deep learning”, Int. J. Archit. Comput., vol. 16, no. 4, pp. 306–327, 2018.

[BRO 21] M. Bronstein, J. Bruna, T. Cohen and P. Velickovic. “Geometric Deep Learning : Grids, Groups, Graphs, Geodesics, and Gauges“, 2021.

[CAP 18] J.F. Capeille, S. Davies, X. Fang, C. Girard et T. le Dantec, “Bien Vivre la Ville : vers Urbanisme Favorable à la santé“, Fondation AIA, Institut CFLD, 2018.

[MAR 19] X.Marsault and F. Torres, “An interactive and generative eco-design tool for architects in the sketch phase”, CISBAT, EPFL, 4-6 september 2019. IOP's Journal of Physics : Conference Series Vol. 1343, november 2019. https://iopscience.iop.org/issue/1742-6596/1343/1