Downstream applications for representation learning in volumetric design tasks

Abstract

Architectural design tasks are essentially sequential in nature. Based on this intuition, here we propose an alternative perspective on data-driven automation in architectural design tasks. Our key idea is to encode the design knowledge from a collection of expert or high-permorning sequential designs and extract useful representations. Later we propose to utilize those learned representations for crucial downstream applications such as design preference evaluation and procedural design generation. Specifically, we demonstrate our idea by leveraging a novel dataset of sequential volumetric designs. Volumetric design, also called massing design, is the first and critical step in professional building design. Many efforts have been made to automatically generate reasonable volumetric designs, but the quality of the generated design solutions varies, and evaluating a design solution requires either a prohibitively comprehensive set of metrics or expensive human expertise. To improve the efficiency and accuracy of volumetric design evaluation, we propose a learning-based approach to create a preference model trained only from high-performing building design solutions. We assume the high-performing designs should have inherent similarities and that we can capture these similarities by extracting features from the high-performing solutions. The preference model can then compare two arbitrarily given volumetric designs. In addition, not only can our model evaluate static solutions but also the design process as the model can take time-series data as input. Finally, we develop an autoregressive model to generate volumetric designs from a partial design input and discuss its capabilities using the novel dataset.

Learning a preference model from high performance design sequences

Results