Multi-point optimization design of high pressure-ratio centrifugal impeller based on machine learning

The traditional optimization method on turbomachinery has the problem as time-consuming and difficult to solve well multi-parameter optimization in a short time. So an accurate surrogate model that is used to estimate the functional relationship between the independent variable and objective value is important to accelerate computationally expensive CFD-based optimization. Many of them had been developed and proven their reliability, such as the Kriging model, Back Propagation Neural Network (BPNN), Artificial Neural Network (ANN) and Support Vector Regression (SVR), etc. Because SVR is based on the principle of structural risk minimization, it has advantages in dealing with a small database, high dimensional and non-linear problems. The reliability of SVR is depended on its kernel parameter and penalty factor, and it can be improved by getting optimal parameters with some optimization algorithms. In this paper, a machine learning model based on SVR combined with a multi-point genetic algorithm (MPGA) is applied to the optimization of a centrifugal impeller with 41 parameters. The optimization objectives are to maximize efficiency at design point and pressure-ratio at near stall point and to minimize the variation of choked mass flow. Results show that (1) time costs reduced significantly. (2) The maximum efficiency increases by 1.24%. (3) It is verified that the reliability of the SVR-MPGA model for multi parameters optimization by comparing to the results of the traditional optimization method-Design 3D The efforts of this study expand the application of machine learning and provide an idea for multi-point optimization of turbomachinery as well.