An alternative approach to surrogate averaging for a centrifugal impeller shape optimisation

A major concern regarding surrogate model-based design optimisation is the modelling fidelity of the approximation functions. This issue is efficaciously addressed by utilising multiple surrogates based on the same data to offer approximations from alternative modelling perspective. The approach is introduced to optimise performance of a centrifugal impeller as a case study. The basic surrogate models adopted here include response surface approximation, radial basis neural network, Kriging, support vector machine and Shepard method. A weighted average and a new weighted average surrogate models were constructed from the basic surrogates and implemented in the present problem. A hybrid genetic algorithm was used to explore the optimal points. Design variables from the impeller inlet and exit blade angles were selected and design of experiments were used to select the sample points from the design space. The aim of the optimisation was to maximise hydraulic efficiency and head generated. The optimised centrifugal impeller yielded lower losses by altering the vane angles. It was found that the most accurate surrogate did not always lead to the best design. This manifested that using multiple surrogates can improve the robustness of the optimisation at a minimal computational cost.