Fast-running metamodels that approximate multivariate input/output relationships of time-consuming physics-based computer simulations (PBCS) enable effective probabilistic analyses of the PBCS outputs under input uncertainties. The probabilistic measures of the simulation outputs can support uncertainty statements about PBCS predictions. In this paper, a general multivariate metamodeling strategy driven by sample cross-validation error metrics will be discussed. A localized regression method using the cross-validated moving least squares (CVMLS) method and an interpolation method using the cross-validated radial basis functions (CVRBF) are developed. A simple example will be presented to illustrate the effectiveness of CVMLS in capturing the highly nonlinear inputs/output relationship.
展开▼
