A Robust Procedure for Convergent Nonparametric Multivariate Metamodel Design
Fast-running metamodels (surrogates or response surfaces) that approximate multivariate input/output relationships of time-consuming CAE simulations facilitate effective design trade-offs and optimizations in the vehicle development process. While the cross-validated nonparametric metamodeling methods are capable of capturing the highly nonlinear input/output relationships, it is crucial to ensure the adequacy of the metamodel error estimates. Moreover, in order to circumvent the so-called curse-of-dimensionality in constructing any nonlinear multivariate metamodels from a realistic number of expensive simulations, it is necessary to reliably eliminate insignificant inputs and consequently reduce the metamodel prediction error by focusing on major contributors. This paper presents a robust data-adaptive nonparametric metamodeling procedure that combines a convergent variable screening process with a robust 2-level error assessment strategy to achieve better metamodel accuracy.