Aljaž Skerlavaj

Centrifugal pumps are widely used in industrial applications. Double-suction (also called: double-entry) centrifugal pumps allow transportation of greater flow rates than single-entry pumps because they are less prone to cavitation problems, and often reach higher efficiency. Another advantage is counter-balancing of axial hydraulic forces due to double-entry design. Usually, hydraulic turbomachines are designed by performing modifications of a nearly-suitable and experimentally validated baseline geometry.

The modifications are either performed directly in a CAD software by the designer, or by modifying input variables in in-house codes using direct or inverse singularity methods. Nowadays, modified geometries are compared by the usage of Computational Fluid Dynamics (CFD) simulations. Only the best design is then usually experimentally tested and manufactured.

There are numerous benefits of using optimization techniques to perform modifications in automatized way instead of using traditional human-based "trial-and-error" technique. First of all, no occasional hard-to-identify human-based errors are present, which guarantees the same design and CFD procedure for each design. Secondly, after the (large) initial effort is performed to include the design creation in optimization loop, the human interaction is reduced and a lot of variants can be tested, which in turn allows a designer to understand the effect of design variables on the objective variable (e.g., pump efficiency) much better. Third, the design space is explored in a more systematic way, giving the opportunity to identify non-trivial design candidates. In this study, the objective of the optimization of a double-suction pump with the specific speed nq equal to 62 (specified per impeller side) is the maximization of its hydraulic efficiency.

During the optimization, only the impeller geometry is modified while the rest of the geometry is fixed. The optimization is performed, by means of the modeFRONTIER optimization platform, in steps. At first, by means of a DOE (Design of Experiments) strategy, the design space is explored, using a parameterized CAD representation of the pump. Suitable metamodels (surrogates or Response Surfaces), which represent an economical alternative to the more expensive 3D CFD model, are built and tested.

Among different metamodels, the evolutionary design, radial basis function and the stepwise regression models seem to be the most promising ones. Finally, the stepwise regression model, trained on a set of 200 designs and constructed with only five the most influential input design parameters, was chosen as a potentially applicable metamodel.