The optimization of a centrifugal compressor impeller is a challenge for established strategies and algorithms, as the interactions between the geometric design parameters and the aerodynamic and structural performance of the system are highly complex. Furthermore many geometrically valid designs are unusable in terms of structure mechanics or flow physics. Due to the complex parameter correlations, a simple limitation of the parametric space is no option, as possibly beneficial parameter combinations could be ruled out. To obtain a meaningful optimization result, the complete operation range of the compressor has to be taken into account which adds further complexities in terms of the optimization process and the computational expense. The combination of these issues leads to a complicated optimization scenario.
The aim of the presented work is the reduction of the computational expense required to generate a high quality metamodel for optimization. This goal shall be achieved by the development of a multi-fidelity sampling method. The basic idea is to use preliminary of low-fidelity information from empirical data or fast analytical methods to identify promising regions of the parameter space. Then the samples of the DOE are concentrated in these areas while still maintaining a good coverage of the whole applicable design space. This ensures that no beneficial designs are ruled out which were not recommended by preliminary information. The points of the resulting DOE are computed by CFD and FEA computations and used to generate the metamodel which is used for the optimization. The method is tested by generating a metamodel used for compressor optimization. The results are compared to an optimization using a metamodel based on a conventional Latin hypercube sampling.