Machine Learning for Practical Real-Time Test Campaign Design

Efficient test campaign design is essential for rapidly validating and verifying aerospace systems. This paper proposes an approach utilizing Support Vector Machines (SVMs) to optimize the design of hypersonic test campaigns in real time. SVMs are employed to predict the utility of candidate test configurations based on datasets acquired a priori. The proposed methodology aims to enhance the efficiency of test campaign design by identifying configurations most likely to yield informative data, which has the potential to reduce resource expenditure and accelerate the pace at which a researcher can acquire valuable experimental and computational data. This paper presents the formulation of SVM-based predictive models tailored to the unique requirements of the Boeing/AFOSR Mach-6 Quiet Tunnel (BAM6QT), detailing the feature selection process and model training. Heat transfer over a finite-span compression ramp on a sliced 7 degree half-angle cone is used as a configuration of investigation for SVM training. The adaptive sampling architecture defines meaningful and consistent categories for heat transfer data. SVMs, as well as other sampling techniques, are evaluated on a distribution of subsonic flow separation. The SVM approach to test configuration selection is shown to outperform most alternative methods for the subsonic tests, though more datasets will be needed to assert generalizability. Prediction of experimental heat transfer by multi-fidelity Kriging with a corresponding CFD dataset is shown to reduce prediction error of mean heat transfer.