This study proposes a framework for data-driven generative design incorporating patent information. Although data-driven generative design provides diverse alternatives by using a deep generative model that generates new alternatives based on automatically discovered latent features, these alternatives depend on the formulation of a design problem. Since not all evaluation criteria can be directly expressed as quantitative metrics, geometric constraints play a crucial role in the early stages of design. However, the correlations between geometric features and evaluation criteria are typically implicit. In contrast, these correlations are obviously embedded in the design results. The proposed framework extracts the relationships between geometric features and evaluation criteria from categories of alternatives collected from patent documents using a concept identification approach. First, a dataset is constructed using drawings and textual descriptions from patent documents. Each alternative in the dataset is represented as an image and a corresponding text. Second, the shapes of the alternatives are converted into combinations of components using image processing techniques to extract geometric features. Third, the alternatives are categorized by clustering based on the geometric features and labeling informed by textual descriptions. Finally, data-driven generative design explores promising alternatives within the selected category. This study verifies the proposed framework by applying it to design problems of tire tread pattern, which involves a high degree of design and diverse evaluation criteria.
TSUMOTO et al. (Wed,) studied this question.