CasaMatic: Performance-Based Building Generation via Conditional Autoencoders

Abstract

This study demonstrates the integration of generative artificial intelligence and building energy simulations to enhance design exploration through inverse design. Using a Conditional Autoencoder (CAE), a type of neural networks, our workflow reverses the traditional forward design process, allowing users to generate building geometries based on design conditions ranging from energy efficiency to different climate scenarios. We used a parametric model to generate the dataset, on which a CAE, comprising an encoder and a decoder, was trained and validated through the AI-extended Design (AIXD) toolbox. During inference, the decoder can be used to generate diverse design options based on requested targets (e.g. low cooling loads for a specific floor-area), while the encoder works as a surrogate model to estimate the performance gap of suggested designs. We applied the model to different case studies in Zürich, successfully generating diverse design alternatives under different boundary conditions. The trained model consistently met heating and cooling targets across different requests such as varying warming scenarios and plot configurations. In out-of-distribution and highly dimensional requests, generation capacity was degraded but could still provide insightful designs. Key challenges ahead include achieving higher degrees of control in querying inverse designs with important architectural and urban constraints, thus avoiding the generation of seemingly random designs, and expanding relevant dimensions such as material compositions and environmental impacts. This study shows the potential of generative approaches to tackle the design of energy efficient buildings in challenging and varied scenarios, underlining the potential of deep learning and AI in the field of architecture.