Empirical Models for Non-Deterministic Simulation-Based Robust Design

We propose a method for metamodeling non-deterministic computer intensive simulations for use in robust design. Generalized linear models for mean responses and heteroscadastic response variances are iteratively estimated in an integrated manner. Estimators that may be used for predicting the mean and variance models are introduced and metamodels of variance are developed. The usefulness of this metamodeling approach in efficient uncertainty analyses of non-deterministic, computationally-intensive simulation models for robust design methods is illustrated with the example of the design of a linear cellular alloy heat exchanger with randomly distributed cracks in the cell walls.