To establish metamodels for the multi-material structure design problems, the material selection of each component is considered as a categorical design variable. One challenging task is to establish an accurate mixed-variable metamodel. It is critical to reduce the prediction error of the mixed-variable metamodel in order to achieve a feasible design with superior performance in the metamodel-based optimization. This paper investigates two different strategies of mixed-variable metamodeling: “feature separating” strategy and “all-in-one” strategy. A supervised learning-aided method is proposed to improve the “feature separating” metamodels. The proposed method is compared with several existing mixed-variable metamodeling methods on three engineering benchmark problems to understand their relative merits. These methods include Neural Network (NN) regression, Classification and Regression Tree (CART) and Gaussian Process (GP). A new Polynomial Coefficient Metric is developed to quantify the adequacy of training data. This study provides insight and guidance for establishing proper metamodels on multi-material structural design problems.

This content is only available via PDF.
You do not currently have access to this content.