Machine learning based design optimization of centrifugal impellers

Abstract

Big data and machine learning are developing rapidly, and their applications in the aerodynamic design of centrifugal impellers and other turbomachinery have attracted wide attention. In this paper, centrifugal impellers with large flow coefficient (0.18–0.22) are taken as research objects. Firstly, through one-dimensional design and optimization, main one-dimensional geometric parameters of those centrifugal impellers are obtained. Subsequently, hundreds of samples of centrifugal impellers are obtained by using an in-house parameterization program and Latin hypercube sampling method. The NUMECA software is used for CFD calculations to build a sample library of centrifugal impellers. Then, applying the artificial neural network (ANN) to deal with the data in the sample library, a nonlinear model between the flow coefficients, the geometric parameters of these centrifugal impellers and the aerodynamic performance is constructed, which can replace CFD calculations. Lastly with the help of the multi-objective genetic algorithm, a global optimization is carried out to fulfull a rapid design optimization for centrifugal impellers with flow coefficients in the range of 0.18–0.22. Three examples provided in the paper show that the design and optimization method described above is faster and more reliable compared with the traditional design method. This method provides a new way for the rapid design of centrifugal impellers.