Optimal design of computer experiments for metamodel generation using I-OPT (TM)

We present a new and unique software capability for finding statistical opt imal designs of deterministic experiments on continuous cuboidal regions. T he objective function for the design optimization is the minimization of th e expected integrated mean squared error of prediction of the metamodel tha t will be found, subsequent to the running of the computer simulations, usi ng the best linear unbiased predictor (BLUP). The assumed response-model fu nction includes an unknown, stochastic term, Z. We prove that this criterio n, which we name I-z-optimality, is equivalent to I-z-optimality for non-de terministic experiments, in the limit of zero correlations among the Z's fo r different inputs. An example is presented of metamodel generation for a m icromachined-silicon flow sensor. The I-z-optimal set of inputs is found, f inite-element (FE) simulations run, and the metamodel generated using a BLU P fit. The method is compared to other approaches. I-z-optimality, coupled with BLUP fitting, provides a highly efficient means of non-parametric meta model generation. I-z-optimal design searching and BLUP fitting are new opt ions of the I-OPT(TM)d program that is available on the World-Wide Web at U RL http://www-personal.engin.umich.edu/similar to crary/iopt.