Early Phase Performance Driven Design Assistance Using Generative Models

Abstract

Form-finding in the current performance-driven design methodology of architectural design is typically formulated as a design optimization problem. Although effective in engineering or late-stage design problems, optimization is not suitable for the exploratory design phase due to the time intensity and cognitive load associated with the processes involved in the formulation and solution of optimization problems. The iterative, diverging nature of early-phase design is incompatible with the i) cognitive load of parametric modeling and its limited affordances for conceptual changes, ii) time and resource intensity of simulations, iii) interpretability of optimization results. This thesis suggests a framework for generating optimal performance geometries within an intuitive and interactive modeling environment in real-time. The framework includes the preparation of a synthetic dataset, modeling its probability distribution using generative models, and sampling the learned distribution under given constraints. The several components are elaborated through a case study of building form optimization for passive solar gain in Boston, MA, for a wide range of plot shapes and surroundings. Apart from the overall framework, this thesis contributes a series of methods that enable its implementation. A geometric system of orientable cuboids is introduced as a generalizable, granular modeling vocabulary. A method for efficient boundary condition sampling is suggested for the dataset preparation. A Variational Autoencoder (VAE) is extended for performance-aware geometry generation using performance-related loss functions. A series of techniques inspired by the data-imputation literature is introduced to generate optimal geometries under constraints. Last, a prototype is presented that demonstrates the abilities of a system based on the suggested framework.