Abstract
Statistical models called metamodels empirically capture the input-output relationship of the analyses for evaluating the objective functions and constrains for the engineering optimal design. They are utilized for three reasons. The first of which is to obtain global obtain global behavior of the original functions that have complex local noises. The second is to shorten optimization calculation time. The third is to enable using the different codes for conjugated analysis. However, the right optimization will not be performed if the approximation has large error. Many metamodeling techniques are proposed and used for optimal designs. In this paper, 7 metamodeling techniques and 4 sampling techniques are compared and evaluated to select the best metamodeling techniques and the best sampling techniques that give the least errors. As a result, it is demonstrated that Kriging, radial basis functions interpolation and radial basis function network give less errors and that the new sampling technique that is proposed in this paper is more effective than the conventional space filling method.
