Resampling Strategies to Improve Surrogate Model-based Uncertainty Quantification - Application to LES of LS89

TL;DR

This paper introduces two resampling strategies to enhance Gaussian Process surrogate models for uncertainty quantification, demonstrated on high-dimensional inputs and applied to LES of the LS89 turbine blade cascade.

Contribution

The paper proposes novel resampling techniques for parameter space to improve surrogate model accuracy in high-dimensional UQ problems.

Findings

Resampling strategies improve surrogate model predictive quality.

Methods are effective for high-dimensional analytical input functions.

Successful application to LES of LS89 turbine blade cascade.

Abstract

Uncertainty Quantification (UQ) is receiving more and more attention for engineering applications in particular from robust optimization. Indeed, running a computer experiment only provides a limited knowledge in terms of uncertainty and variability of the input parameters. These experiments are often computationally expensive and surrogate models can be constructed to address this issue. The outcome of a UQ study is in this case directly correlated to the surrogate's quality. Thus, attention must be devoted to the Design of Experiments (DoE) to retrieve as much information as possible. This work presents two new strategies for parameter space resampling to improve a Gaussian Process (GP) surrogate model. These techniques indeed show an improvement of the predictive quality of the model with high dimensional analytical input functions. Finally, the methods are successfully applied to a turbine blade Large Eddy Simulation application: the aerothermal flow around the LS89 blade cascade.