A Hybrid Method for Surrogate Model Updating in Engineering Design Optimization

Abstract

This paper addresses the critical issue of effectiveness, efficiency, and reliability in simulation-based design optimization under surrogate model uncertainty. Specifically, it presents a novel method to build surrogate models iteratively with sufficient fidelity for accurately capturing global optimal design solutions at a minimal cost. The salient feature of the proposed method lies in its unique preference of focusing necessarily high fidelity at potential global optimal regions of surrogate models. The proposed method is the synergic integration of the multiple preference point method, which updates surrogate model at current local optimal points predicted with data-mining techniques in genetic algorithm setup, and the maximum variance point method, which updates surrogate model at the point associated with the maximum prediction variance. Through illustrative comparison studies on thirty different optimization scenarios derived from 15 different test functions, the proposed method demonstrates the tangible reliability advancement. The experimental results indicate that the proposed method can be a reliable updating method in surrogate-model-based design optimization for efficiently locating the global optimal point/points in various kinds of optimization scenarios featured by single/multiple global optimal point/points that may exist at the corners of design space, inside design space, or on the boundaries of design space.