Investigation of novel approaches for aerodynamic data fusion

Kurzfassung

Modern industrial applications require reliable and accurate aerodynamic data for design and optimization. This data is generally produced using CFD simulations and wind tunnel testing. Although these approaches offer significant individual benefits, they also have certain limitations. CFD fails to yield accurate solutions towards the edge of the envelope, whereas the wind tunnel experimental data offers data only at specific sensor locations. Data fusion techniques combine the individual strengths of these data sources to deliver accurate and reliable data. POD-based data fusion techniques like Gappy POD and regularized Gappy POD are well-established and widely used in various studies. These techniques compute the data fusion result via a least-square fit in the POD subspace. Recently shallow artificial neural networks have also been used in data fusion techniques to reconstruct the flow solution. This thesis proposes an alternative data fusion approach called Gappy ANN and compares it with Gappy POD. The idea of Gappy ANN is to replace the POD subspace with a solution space generated via a shallow artificial neural network. The advantage of this approach is that knowledge of the sensor positions can be directly considered when creating the solution space. This thesis demonstrates the performance and robustness of Gappy POD and Gappy ANN on an aerodynamic test case fusing highquality experimental and numerical data. Gappy POD performs better in reconstructing the flow solution than Gappy ANN, showing only minimal errors. To improve their prediction accuracy, DEIM-based sensor placement strategies are applied to the POD reduced space and ANN solution space to obtain the new optimal sensor locations. Gappy POD and Gappy ANN solutions are recomputed using these new locations. The results for both experimental and DEIM-based indices are analyzed to demonstrate the significance of the DEIM-based algorithm for the Gappy approaches in improving prediction accuracy.