Metamodels are extensively utilized to replace the computational-cost simulation models in calibration, performance analysis and design optimization. Various types of experiment design methods have been proposed for the metamodeling. In this paper, an adaptive sequential sampling method based on Delaunay triangulation is proposed. A support distance criterion is designed to measure the space filling property of candidate design points and a local linear appropriation approach is provided to estimate the prediction error. Furthermore, the metamodeling process based on the proposed adaptive sampling method is discussed. Finally, a systematic comparison is conducted to evaluate the effectiveness of the proposed adaptive sequential sampling method (ASED) and two types of metamodels (Gaussian process model and adaptive regression splines). The numerical experiments indicate that the GP model constructed with ASED achieves the best prediction performance.
展开▼
