Pump Suction Shape Optimization Using a Parallel Stochastic Radial Basis Function Method

This paper deals with a shape optimization of pump suction, with the objective of improving the pump performance. A combination of ANSYS CFX software tools and a surrogate-based, so-called multistart local metric stochastic RBF (MLMSRBF) method for the global optimization of “expensive black-box functions” is employed. The shape of the suction is driven by 18 geometric parameters, and the cost functional is based on the CFD results. The practical aspects of assembling and evaluating a parametric CFD model, including mesh independence study, are shown. After initial design of experiment evaluation, a response surface model is created and used for generating new sample points for the expensive CFD evaluation. Then, the whole process is repeated as long as necessary. A parallel version of the method is used, with necessary modifications for dealing with CFD-specific problems, such as failed designs and uncertainty of computational times. Both steady-state and transient models are used for the optimization, each with a different objective function. The resulting designs are then compared with the original geometry, using a complete model of the pump and fully-transient simulation. Both hydraulic characteristics and multiphase cavitational simulations are considered for the comparison. At the end, the results and challenges of using these methods for CFD-driven shape optimization are discussed.