Early-Phase Performance-Driven Design Using Generative Models

Current performance-driven building design methods are not widely adopted outside the research field for several reasons that make them difficult to integrate into a typical process. In early phase, in particular, time intensity and cognitive load associated with optimization form parametrization incompatible exploration, which requires quick iteration. This introduces novel method geometry generation can afford interaction directly 3d modeling environment, eliminating need explicit parametrization, is multiple orders faster than equivalent optimization. The uses Machine Learning techniques train generative model offline. learns distribution of optimal performing geometries their simulation contexts based on dataset addresses performance(s) interest. By navigating model’s latent space, desired characteristics be quickly generated. A case study presented, demonstrating synthetic use Variational Autoencoder (VAE) as solar gain. results show VAE-generated perform average at least well optimized ones, suggesting introduced shows feasible path towards more intuitive interactive early-phase assistance.