Deep concept identification for generative design

Ryo Tsumoto

doi:10.1016/j.aei.2025.103354

Return

Deep concept identification for generative design

2025-05-01

Ryo Tsumoto

Kentaro Yaji

Yutaka Nomaguchi

DOI：10.1016/j.aei.2025.103354

Original

Save

View PDF

Abstract

En 中文

Generative design techniques have become sophisticated methods for generating diverse alternatives by incorporating topology optimization with artificial intelligence techniques. As their diversity increases, the cognitive burden on designers in selecting the most appropriate alternatives also increases. The concept identification approach, which finds various categories of entities, is expected to be effective for systematically interpreting their diversity. However, conventional concept identification approaches cannot provide meaningful categories when their geometric properties face high-dimensionality. To address this challenge, this study proposes a new concept identification framework for generative design using deep learning (DL) techniques. One of the key abilities of DL is the automatic learning of effective representations of a specific task. This study first outlines the key points of concept identification based on the general design theory, then proposes a basic framework that consists of generating diverse alternatives using a generative design technique, clustering the alternatives into several categories using a DL technique, and arranging these categories into design concepts using a classification model. This study demonstrates its fundamental capabilities by implementing variational deep embedding, a generative and clustering model based on the DL paradigm, and logistic regression as a classification model. Its implementation is applied to a simplified design problem of a two-dimensional bridge structure as a case study. The proposed deep concept identification framework can systematically identify meaningful categories of diverse alternatives, while it still requires designer cognition in several steps because of the gap between the data-driven approach and the nature of concept identification.

Keywords:

Concept identification

Design concept

Generative design

Topology optimization

Representation learning

Clustering