A Parallel Multi-objective Efficient Global Optimization Method and Its Application of Multi-stage Axial-Flow Compressor Optimization Design

The essence of optimization design of a multi-stage axial flow compressor is to solve an expensive black-box multi-objective problem. The multi-objective efficient global optimization has been demonstrated as a promising alternative to solve such expensive multi-objective problems. It sequentially samples the design space to seek for the Pareto solutions with a minimum number of sample points. With access to parallel computation, it is wise to add multiple infill points per optimization iteration to make optimization parallelizable to shorten the total design process further. In this paper, the parallel version of the expected improvement of modified hypervolume (EIMHV) criterion is developed following the multiple good local optima strategy. An empirical parameter-free exclusion radius function is proposed and combined with the EIMHV criterion to easy the search of the multiple local optima. The proposed method is used to solve analytic benchmark problems firstly. It is found that the parallel method with moderate number of infill points could achieve significant speedup without noteworthy sacrifice to the optimization gain. While, parallel method with a large number of infill points cannot provide further acceleration to the optimal search but deteriorate the optimization performance. By incorporating an efficient Kriging modeling method for high-dimension problem and the imputation strategy to deal with simulation failures to the proposed method, the optimization design of a 3-stage axial flow compressor with 2 objectives and 144 design variables is solved. Significant improved objective values and convergence speed demonstrated the engineering practicality of the proposed method.