Data mining-assisted multi-objective optimization of blade angle distributions for efficiency and stability enhancement of a free-form centrifugal impeller

The design optimization process for a centrifugal compressor faces significant trade-offs among efficiency, pressure ratio, and stable operating range. To address this challenge, a data mining assisted-optimization method combining machine learning modeling and non-dominated sorting genetic algorithm III (NSGA-III) has been proposed to enhance efficiency and pressure ratio while considering the expansion of operational stability. A validated CFD numerical model was established for data generation. A Light-GBM model with Bayesian hyperparameter tuning and a deep feedforward neural network (DNN) model were constructed, of which the better one was selected for performance prediction. Data mining techniques, including Sobol sensitivity analysis and SHapley Additive exPlanations (SHAP), were employed to identify the key sensitive parameters and their interactions. Finally, an NSGA-III-based multi-objective optimization was conducted to determine the optimal geometry of the free-form surface impeller. Results showed that the blade angle distributions at the leading edge of mid-span and shroud profile were the most sensitive parameters affecting isentropic efficiency at the design point, with their recommended ranges identified. The optimized impeller has an increase in total pressure ratio of 2.98 % and isentropic efficiency of 0.67 % at the design point, especially exhibiting excellent aerodynamic performance at higher mass flow rate conditions. The maximum improvements of total pressure ratio and isentropic efficiency were 12.91 % and 9.53 %, respectively, corresponding to the near choke mass flow rate at the design speed. Additionally, operating stability was expanded with a 1.22 % increase in total pressure ratio at near surge mass flow rate at the design speed, which contributed to a delay in the onset of surge. Regarding the computational cost, the data mining-assisted optimization can run faster with its required time saving 12.3 %.