Multifidelity methods for multidisciplinary design under uncertainty

Abstract

For computational design and analysis tasks, scientists and engineers often have available many different simulation models. The output of each model has an associated uncertainty that is a result of the modeling process. This uncertainty is referred to as model discrepancy and is defined as the deviation of the model output relative to the "true" physical value. The design process typically begins with computationally inexpensive, lower fidelity models and advances to the higher fidelity models as knowledge of the design space is acquired. Previous research has developed a Bayesian-based multidisciplinary design optimization (BMDO) framework for conducting multifidelity design with uncertainty. Fidelity level is associated with the magnitude of model discrepancy. Model selection is determined by apportioning design uncertainty to the disciplines to identify key contributors. As fidelity level increases, information from the lower fidelity models is used to complement the higher fidelity results through information fusion instead of being discarded, a more traditional approach in multifidelity optimization. This research expands on the previously developed BMDO framework by investigating the effects of interdisciplinary coupling and model correlation on the design process. Uncertainty in the coupling variables is introduced to the BMDO framework. Multifidelity models tend to be founded on similar underlying physics and numerical methods. As a result, the model output from different fidelities may exhibit non-negligible correlation. This research demonstrates that exclusion of model correlation and uncertainty due to interdisciplinary coupling may result in underestimates of the uncertainty in design quantities of interest.