In this work, the polynomial chaos-based Cokriging (PC-Cokriging) is applied to a benchmark aerodynamic design optimization problem. The aim is to perform fast design optimization using this multifidelity metamodel. Multifidelity metamodels use information at multiple levels of fidelity to make accurate and fast predictions. Higher amount of lower fidelity data can provide important information on the trends to a limited amount of high-fidelity (HF) data. The PC-Cokriging metamodel is a multivariate version of the polynomial chaos-based Kriging (PC-Kriging) metamodel and its construction is similar to Cokriging. It combines the advantages of the interpolation-based Kriging metamodel and the regression-based polynomial chaos expansions (PCE). In the work the PC-Cokriging model is compared to other metamodels namely PCE, Kriging, PC-Kriging and Cokriging. These metamodel are first compared in terms of global accuracy, measured by root mean squared error (RMSE) and normalized RMSE (NRMSE) for different sample sets, each with an increasing number of HF samples. These metamodels are then used to find the optimum. Once the optimum design is found computational fluid dynamics (CFD) simulations are rerun and the results are compared to each other. In this study a drag reduction of 73.1 counts was achieved. The multifidelity metamodels required 19 HF samples along with 1,055 low-fidelity to converge to the optimum drag value of 129 counts, while the single fidelity models required 155 HF samples to do the same.
展开▼
