Multi-fidelity gaussian process for uncertainty quantification in aerodynamic analysis

Abstract

This research contributes to the AIAA FD UQ Discussion Group’s Challenge Problem, with an aim to evaluate aerodynamic coefficients and their total uncertainty for an NACA 2412 airfoil. In this research, the problem is tackled with Multi-fidelity Gaussian Process (MFGP), where experimental data and XFOIL predictions are considered as two levels of fidelity. The total uncertainty is quantified as the MFGP’s predicted variance. Three training approaches are implemented and tested repeatedly. According to the results, MFGP can effectively capture the nominal values of the aerodynamic coefficients with very sparse experimental data. However, the total uncertainty is underestimated. This is potentially due to the limitation of the maximum likelihood method for estimating the MFGP’s hyperparameters.