Optimal surrogate modelling approaches for combining experimental and computational fluid dynamics data sets

In this study multi-fidelity surrogate modelling for combining data sets of wind tunnel experiments and computations is examined, dealing with different types of errors. Co- kriging regression is constructed with the low-fidelity sample data of the computations and the high-fidelity data of the wind tunnel experiments, and is compared with co-kriging and polynomial response surface approaches. Face-centred central composite design is used to obtain the high-fidelity sample data for the co-kriging and co-kriging regression, where a blocking method is used to prevent systematic error between block boundaries. A randomisation method is for the wind tunnel experiments to reduce systematic error. Co-kriging regression has the potential to reduce the effect of systematic error working with randomisation method. The test case of a race car wing in ground effect is used here, and shows that while the polynomial response surface can not indicate a local optimum, the co-kriging and co-kriging regression do identify the twin optima that can be explored in more detail by adding sample points. The co-kriging regression shows a lower root mean square error compared to the other approximations. For assessing the confidence of surrogate models, the combined uncertainty of the approximations is shown, comprising the modelling uncertainty and the sample data uncertainty.