Prediction of Flow Separations in Turbine Center Frames Using Support Vector Machines and Random Forests

The reduction of overall fuel consumption has become one of the main targets in developing new aero engines, where an increasing bypass ratio has proven to be an efficient way to achieve this goal. For widely spread two-spool configurations, this directly impacts the turbine center frame (TCF), located between the high- and low-pressure turbine. For higher bypass ratios, the component has to deal with bigger radial offsets and stronger curvatures, making the channel prone to flow separation and all the undesirable side effects. A large number of parameters and nonlinear interactions between them make it very challenging to define limits within the parametric space beyond which flow separations may occur. In order to investigate this behavior, a large data set of CFD calculations is created, comprising multiple geometrical and boundary-related parameters. The data set is used to train random forest classifiers (RFC) and support vector classifiers (SVC) to predict the appearance of flow separations for given parametric combinations. Both algorithms are tested on a separate data set and are compared in terms of accuracy. Further investigations are dedicated to the identification of the most crucial parameters. Finally, a linear SVC is used to derive an easily interpretable limit in parameter space, in order to avoid flow separations.