Fast-running metamodels that approximate multivariaterninput/output relationships of time-consuming physics-basedrncomputer simulations (PBCS) enable effective probabilisticrnanalyses of the PBCS outputs under input uncertainties. Thernprobabilistic measures of the simulation outputs can supportrnuncertainty statements about PBCS predictions. In thisrnpaper, a general multivariate metamodeling strategy drivenrnby sample cross-validation error metrics will be discussed. Arnlocalized regression method using the cross-validatedrnmoving least squares (CVMLS) method and an interpolationrnmethod using the cross-validated radial basis functionsrn(CVRBF) are developed. A simple example will be presentedrnto illustrate the effectiveness of CVMLS in capturing thernhighly nonlinear inputs/output relationship.
展开▼