Deep shape optimization with limited dataset: leveraging implicit neural representation

Abstract

Generative AI that learn data to generate new shapes of design have a profound impact on mechanical design. Above all, the generative AI model can automatically select features and map them into the latent space without the help of experts to parameterize parameterization, which is the most bottleneck in the parameter-based shape optimization methodology. However, the fact that the AI methodologies currently being studied require a lot of data to be learned about the same class and still operate in a high level latent space from an optimization perspective is considered a limitation in the field of mechanical engineering. This study aims to propose a methodology that extracts the characteristics of a small number of data with only relatively small latent space. Using car and wheel data, this study verified the robustness and versatility of the methodology by conducting experiments that traditional shape optimization methodologies could not perform.