Sequential Metamodeling in Engineering Design

Abstract

Statistical metamodels (surrogate models) are usually developed to replace complex, time-consuming simulation programs or expensive physical experiments and facilitate fast and accurate analysis. Various types of metamodeling methods have been used in the literature, each having its strong and weak points. The choice of the appropriate type of metamodel is important in design. To save time and money spent on expensive experiments, it is also very important to choose a method for sequential experimental design and metamodel development in which data points are identified along a design timeline to help obtain maximum information. Based on the observations in this paper and previous studies, a framework for sequential metamodeling is proposed, in which kriging and MARS metamodels are used to help designers develop appropriate metamodels with limited resources. The method of Sequential Exploratory Experimental Design (SEED) is improved through the utilization of kriging and MARS metamodels.