Toward Automatic Floor Plan Generation through Mathematical Optimization
In the early stages of architectural design, a floor plan must satisfy many conditions at once, including site constraints, legal requirements, room sizes, and relationships among spaces. Because these requirements interact in complex ways, developing a viable plan often demands substantial time and repeated trial and error. This project explores how mathematical optimization can support early-stage design decisions by automatically generating floor plan proposals in collaboration with industry.
Our aim is different from that of typical generative AI systems that interpret conditions loosely and produce plausible-looking results. Instead, we explicitly encode conditions such as site shape, building coverage ratio, floor area ratio, required rooms, zoning relationships, and spatial connectivity as constraints, and then search for plans that are actually feasible. To do this, we model a plan as a set of grid-based units assigned to different functions, while continuously improving the computational method so that it can handle larger and more realistic problems.
As a prototype, we are using conditions similar to those found in architectural licensing exam drafting tasks and developing an algorithm that can automatically produce multiple floor plan options. Improvements in computational performance are making it possible to solve problems in far shorter times than before, even in cases that were previously impractical. Rather than replacing architects, this research seeks to create an environment in which designers can organize complex requirements and make decisions more quickly and more confidently.
