Machine learning and genetic algorithms for conformal geometries in design support systems

To optimise both building designs and their underlying design processes, design support systems exist. For domain specific analyses, these systems benefit from a conformal (CF) representation for the Building Spatial Design (BSD). In a conformal representation, for all entities: the vertices of an entity are, if intersecting another entity, only allowed to coincide with this other entity's vertices. This paper presents research on whether Machine Learning (ML) and Genetic Algorithms (GA) can be used to obtain a conformal geometry for BSDs. For ML, a neural network is trained to learn the complex relation between BSDs and their conformal representations. GAs are first used to find all quad-hexahedrons in the search space, then to find sets of quad-hexahedrons that form the conformal design. A trained ML model does provide outcomes, but not very useful, even with encoding the configuration type of the design. Differently, the GA finds conformal designs for many instances, even for non-orthogonal designs.