Variational deep embedding mines concepts from comprehensive optimal designs

This paper aims at exploring the possibility of mining design concepts from superior design samples, which are comprehensively generated by computa tional optimization, by means of the variational deep embedding (VaDE), which is a technique of deep learning with dimensionality reduction and clustering. The relationships between concepts and entities have been a crucial question in design research. This paper reports on an experiment that a VaDE encodes a set of con ceptual designs of a bridge structure generated by a topology optimization tech nique into a set of clusters that are represented as Gaussian distributions in the latent space. In this experiment, it is confirmed that each cluster corresponds to a repre sentative design concept and that its associated subspace corresponds to a classi f ication of design possibilities. Following such a result, the promises and challenges of concept mining are discussed.