Machine-learnt topology of complex tip geometries in gas turbine rotors

The work presents a data driven based strategy to develop a new statistical model of complex tip shape for high-pressure turbine stages exploiting an existing dataset of optimized squealer-like rotor tips. Using the exploratory data analysis (EDA), a set of statistical methods were used to improve the quality of previous CFD-based optimization dataset, as an aid in reducing outliers, data skewness and avoiding the presence of redundant information. The pre-processed dataset was analyzed by unsupervised learning method in order to gain insight on the correlation between tip geometry and single stage axial turbine performance. Utilizing the Principal Component Analysis (PCA), we developed a new continuous, dimensionality-reduced parametrization which allows overcoming the limitations of discrete topology approaches. The novel statistical shape model, coupled with genetic operators into a NSGA-II optimization strategy, was used to explore the design space of optimal solutions generating new designs to enrich the available Pareto front in terms of aero-thermodynamic performance metrics. Two metamodels for performance prediction, respectively based on Artificial Neural Network (ANN) and Gradient Boosting Regressor (GBR) have been developed in order to guide the Pareto front exploration avoiding the use of computationally intensive CFD simulations. New tip designs were carried out to spread the previous optimal front and, successively, aiming to design individuals able to reduce macroscopic not uniformity of the flow keeping optimal aerodynamic performance.