EFFICIENT UNCERTAINTY QUANTIFICATION WITH GRADIENT-ENHANCED KRIGING: APPLICATIONS IN FSI

Techniques for Uncertainty Quantification (UQ) suffer fromthe 'curse of dimen- sionality': the number of required evaluations of the simul ation code increases rapidly as the number of uncertain variables increases. Fluid-StructureInteraction (FSI) problems can in- volve complex physics as well as a large number of random inpu t variables. The objective of the current work is to mitigate the curse of dimensionality b y including adjoint-based gradient information from the FSI problem. For a FSI problem we increa se the number of random struc- ture variables from1 to 16. We apply a UQ response surface technique known as Kriging, a nd observe the computational effort that is required to obtaina certain target accuracy. When in- cluding gradient information - a technique known as Gradien t-Enhanced Kriging (GEK) - we find a speedup that increases with the number of random variab les. For example, for4 random variables we observe a speedup of3.0, while for16 random variables we observe a speedup of 9.8. We conclude that including gradient information can lead t o significant speedups.