A Surrogate-based Approach for a Comprehensive UQ Analysis in CFD

Abstract

The main objective of this paper is to presents a holistic approach for uncertainty quantification and propagation applied to an aerodynamics engineering problem. This probabilistic prediction is facilitated by combining different techniques generally involved in a UQ analysis, however, baring in mind the limited computational cost and time available to a high-fidelity simulation engineer/researcher. In this study, we use an efficient surrogate-based modeling strategy of the black-box aerodynamics solver to make faster and cheaper evaluations which are in turn (re)used for further analysis. The results obtained using this approach are found to be in accordance with Monte-Carlo estimates using 500,000 samples. Moreover, a multi-fidelity approach which relies on a variable-fidelity surrogate model offers a reasonably accurate surrogate at a significantly lower computational cost. A surrogate-based double-loop quasi Monte-Carlo approach for propagation of mixed uncertainties is found to be fast and efficient. To study their influence on the outputs, in addition to the model input uncertainties, the uncertainties associated with discretization and surrogate model are also propagated. Furthermore, this paper also presents the results from sensitivity analysis followed by a discussion on simulation credibility. All capabilities are demonstrated on the AIAA UQ challenge problem using the DLR Surrogate Modeling for Aero-Data Toolbox in python.