EVALUATION OF DIFFERENT REGRESSION MODELS TUNED WITH EXPERIMENTAL TURBINE CASCADE DATA

In the present work linear and non-linear regression functions have been tuned with an extensive database describing the unsteady aerodynamic efficiency of low-pressure-turbine cascades. The learning strategy has been first defined using a dataset published in a previous work concerning the loss coefficient measured in a large-scale cascade for a large variation of the Reynolds number, the reduced frequency, and the flow coefficient. Linear models have been educated accounting for the Occam's razor parsimony criterion, condensing the effects due to the parameter variation in few predictors. Then, these predictors have been used to generate an extended polynomial-base function, and an ℓ1-norm constrain has been included into the optimization process to promote sparsity. Different non-linear models have been also evaluated, introducing the formulation of Gaussian Processes for regression for different kernel functions. The capabilities of the models here tuned are compared by means of cross-validation global and local criteria. Cross-validated error and leverage distribution have been analyzed. The proper compromise between model accuracy and generalizability is identified as a Pareto front in the space of the cross-validation indicators. In addition, a new variance-based indicator for the identification of the best model among the candidate ones has been introduced and its capability in complementing the cross-validation analysis is here discussed. The learning strategy has been finally replicated adopting an extremely large database, experimentally acquired in the framework of the extensive collaboration between the University of Genova and AvioAero. For confidentiality, results concerning this database are not shown in terms of response surface, but model performances are discussed in order to strengthen the strategy here adopted, that could be useful also to other research groups adopting their own data.