A study on advanced surrogate model-based optimization using auxiliary information

Abstract

Surrogate models, which approximate real-world systems as mathematical models, have been widely used to solve computationally expensive engineering system. However, despite advances in surrogate modeling and increase in computing power, obtaining a sufficient number of data is limited in engineering fields. To solve this issue, this dissertation proposes an integrated advanced surrogate modeling framework using auxiliary information such as gradient, low-fidelity, and nearby prior information. First, a novel adaptive gradient-enhanced Kriging (AGEK) method is proposed to reduce the total time from simulation to surrogate modeling while achieving the required target performances. The main goal of the proposed method is to adaptively exploit the gradient information, ultimately break down the boundary between Kriging and gradient-enhanced Kriging. Second, an efficient multi-fidelity (MF) surrogate framework based on modified MF dataset selection method is proposed to make the most of cheap auxiliary data. The main purpose of the proposed method is to select best surrogate model between single- and multi-fidelity surrogate models by determining whether to use low-fidelity information. Third, a reanalysis-based multi-fidelity (RBMF) surrogate method, which combines the MF surrogate modeling and a structural reanalysis method, is developed to reduce the computational cost for each sample, not the number of samples. The core idea of the RBMF surrogate modeling is to approximately obtain a large number of data based on a small number of exactly calculated data as prior knowledge. Sub-strategies for the proposed three methods are developed to strengthen the performances of each methodology. The superiorities of each proposed method are verified through various numerical examples and real-world engineering problems. Finally, guidelines on the proposed integrated framework are provided.