Deciphering the noisy landscape: Architectural conceptual design space interpretation using disentangled representation learning

Time and resource restrictions limit the architect's design scope. Computational design methods can offer support to overcome these limitations. Design exploration has been a long-established task in computational-aided generative design. However, conventional objective- and performance-based systems have restrictions pertaining to the exploration scope. Without a quasi-global cognition of the conceptual design space, the exploration scope is bound to be limited. This paper is a proposal for an epistemic shift toward the interpretation of conceptual design space per se. This topic receives limited attention in the current literature due to the scarcity of interpretation tools. Using a customized large-scale architectural image database with high-level visual diversity and latent data space coverage, this paper serves as a first attempt to investigate the possibilities of leveraging disentangled representation learning to structurally interpret architectural conceptual design space in both supervised and unsupervised manner. Various schemes of supervised disentanglement are tested, with analytical comparisons indicating discrepant structural traits of different latent spaces. The unsupervised interpretation scheme shows the preliminary capability of automatic feature disentanglement. Our long-term objective is to offer designers a broader spectrum of creative design through innovative design systems.