Design optimization by integrating limited simulation data and shape engineering knowledge with Bayesian optimization (BO-DK4DO)

Abstract

Surrogate models have been widely studied for optimization tasks in the domain of engineering design. However, the expensive and time-consuming simulation cycles needed for complex products always result in limited simulation data, which brings a challenge for building high accuracy surrogate models because of the incomplete information contained in the limited simulation data. Therefore, a method that builds a surrogate model and conducts design optimization by integrating limited simulation data and engineering knowledge through Bayesian optimization (BO-DK4DO) is presented. In this method, the shape engineering knowledge is considered and used as derivative information which is integrated with the limited simulation data with a Gaussian process (GP). Then the GP is updated sequentially by sampling new simulation data and the optimal design solutions are found by maximizing the GP. The aim of BO-DK4DO is to significantly reduce the required number of computer simulations for finding optimal design solutions. The BO-DK4DO is verified by using benchmark functions and an engineering design problem: hot rod rolling. In all scenarios, the BO-DK4DO shows faster convergence rate than the general Bayesian optimization without integrating engineering knowledge, which means the required amount of data is decreased.