Investigation of multi-fidelity data fusion techniques to obtain an aerodynamic database for a generic fighter aircraft

Aerodynamic data sets are an important source for determining the flight performance, stability and control, and flight dynamics of an aircraft. Aerodynamic data sets contain aerodynamic coefficients corresponding to different flight conditions. These aerodynamic coefficients can be obtained from different sources with different fidelity values. The most accurate values are obtained from flight tests, followed by large and small scale wind tests, computational fluid dynamics simulations and numerical solvers to obtain aerodynamic data of different fidelity. High-fidelity data provides accurate results for the interpretation of aircraft parameters, but obtaining high-fidelity data is labor costly and time consuming. With the availability of multi-fidelity data fusion methods, it is possible to use high and low fidelity data. This reduces the need for only high fidelity data and is more efficient, labor and time saving. The data obtained with the data fusion method is ensured to be high fidelity data. In this study, different data fusion methods were used using low and high fidelity data obtained from the generic fighter jet. These methods are co-Kriging, multi-fidelity Gaussian Process Regression and multi-fidelity neural networks. Two different data sources were used to obtain the data. Two different data sources were used to obtain the data. Low fidelity data were obtained using Vortex-Lattice method in OpenVSP program and high fidelity data were obtained using k-ω SST solver in ANSYS Fluent program. Lift, drag and pitch moment coefficients were obtained from both sources in the range of -15 to +15 degrees angle of attack. Test-cases were created using different amounts of data from both data sources and the performances of all three methods were analyzed. It is stated that the predictions of all three methods match with the high fidelity data as the number of high fidelity data increases In the co-Kriging method, it was observed that the prediction performance changed when the locations of the data points were changed, so the co-Kriging method was more sensitive than the other methods. It was determined that the MF-GPR method gives the best prediction results among the other methods, and while the MF-NN model has good prediction ability on flat and near flat slopes, but it needs more high fidelity data points in the data line with high curvature slope.