UNCERTAINTY QUANTIFICATION OF COMPUTATIONAL FLUTTER ESTIMATES OF A COMPRESSOR CASCADE

Aeroelastic instabilities such as flutter can greatly limit the operating range and safety of modern aircraft engines. Current computational methods have a central role in the evaluation of turbomachinery blades stability, but can be affected by errors if the investigated flow conditions break model assumptions or are particularly sensitive to small changes in flow variables. In this paper, a machine learning based method is proposed to quantify the uncertainty of computational aerodynamic damping predictions. The test case employed for this study is a two dimensional compressor cascade, which resembles most of the relevant aeroelastic features of modern fan and compressor blades. A random forest based model is trained and tested to construct a mapping between input features and aerodynamic damping, both obtained from linearised CFD computations. The input features concern simple, physically relevant quantities that are available early on in design stage. The results show that the machine learnt model can produce predictions, by interpolating within the range of input features, with a coefficient of determination R-2 approximate to 0.94. Moreover, the predictions are enhanced with a measure of uncertainty in terms of confidence intervals. The results show that the confidence intervals can accurately portray the sensitivity of aerodynamic damping with respect to the flow variables. Finally, to underline the relevance of such an approach during design, the model is applied to obtain a conservative flutter boundary on a compressor map, providing a safer operating margin.