Early-Phase Performance-Driven Design using Generative Models

TL;DR

This paper presents a fast, interactive method for early-phase building design using generative machine learning models that produce geometries optimized for performance without explicit parametrization.

Contribution

It introduces a novel approach that leverages generative models, specifically VAEs, for performance-driven geometry creation in design, enabling quick iteration and interaction in 3D environments.

Findings

VAE-generated geometries match or outperform optimized ones in solar gain

Method is several orders faster than traditional optimization techniques

Enables intuitive, real-time design exploration

Abstract

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 design process. In the early design phase, in particular, the time-intensity and the cognitive load associated with optimization and form parametrization are incompatible with design exploration, which requires quick iteration. This research introduces a novel method for performance-driven geometry generation that can afford interaction directly in the 3d modeling environment, eliminating the need for explicit parametrization, and is multiple orders faster than the equivalent form optimization. The method uses Machine Learning techniques to train a generative model offline. The generative model learns a distribution of optimal performing geometries and their simulation contexts based on a dataset that addresses the performance(s) of interest. By navigating the generative model's latent space, geometries with the desired characteristics can be quickly generated. A case study is presented, demonstrating the generation of a synthetic dataset and the use of a Variational Autoencoder (VAE) as a generative model for geometries with optimal solar gain. The results show that the VAE-generated geometries perform on average at least as well as the optimized ones, suggesting that the introduced method shows a feasible path towards more intuitive and interactive early-phase performance-driven design assistance.