MACHINE-LEARNING METHODS FOR DATA ANALYSIS AND OPTIMIZATION OF LPT CASCADES

The present thesis is aimed at presenting some advancements and opportunities provided by modern machine-learning (ML) tools for data analysis and optimization in the field of aeroengine Low-Pressure-Turbine (LPT). Specifically, in this work the ML tools led to the definition of: (O1) a proposal for a new profile loss correlation for LPT cascade, (O2) a blade geometry model to be used alongside the loss correlation for the identification of the optimal LPT blade shapes, and (O3) a new sampling strategy for the optimization of future test campaigns based on either experiments or numerical simulations. A vast database concerning LPT blade aerodynamic performance has been experimentally acquired during the work and has been used as foundation for the application of the ML tools. The database includes the results of several low-speed tests carried out in two wind tunnels installed in the Aerodynamic and Turbomachinery Laboratory of the University of Genova for different families of LPT cascades and for many different operating conditions. The tests also account for the unsteady aerodynamic interaction between the stator and the rotor, that has been simulated during the experiments by means of a moving bars system. For accomplishing the objectives of the work, regression tasks have been addressed. Pros and cons of linear, regularized and non-linear regression models have been deeply analyzed. The advantages provided by Gaussian Processes are particularly highlighted in this thesis. Bayesian and Cross-validation criteria have been used for model selection. Especially, a new robust strategy for the identification of the right compromise between model accuracy and generalizability has been defined and presented in this work. It led to the definition of the new profile loss correlation (O1), which overcomes existing ones because it accounts for the unsteady effects, and because it is dynamic, i.e., easily updatable with new data every time it is deemed necessary. The new correlation led to the identification of a new optimal LPT cascade, the results of which allows the validation of the model and provides further enlargements. ML tools for dimensionality-reduction have also been used to analyze high dimensional data. The Proper Orthogonal Decomposition (POD) has been moved from the classical fluid dynamic framework toward unconventional applications for modelling loss coefficient distributions and interpreting blade geometries. A POD-based geometry model (O2) able to generate LPT blade shapes from a set of few geometrical parameters has been defined. It can be used alongside the loss correlation to quickly provide optimal interpolated blade shapes in the LPT design regions that are expected to minimize losses. Moreover, a POD-based sampling strategy (O3) has been proposed for the identification of the tests that are strictly needed for modelling the cascade performance. Such procedure may be useful for optimizing new campaigns of cascade tests, and more in general to observe the response of a new system with respect to the parameter variation with a coarse grid sampling the design space. The results here obtained constitute a direct support for LPT aero design, while the used ML procedures can be extended to other engineering applications.