Development, Validation and Application of an Optimization Scheme for Impellers of Centrifugal Fans Using CFD-Trained Metamodels

Automatized Aerodynamic Design via Optimization

Cost-effective optimization of centrifugal impellers requires quick and reliable methods to evaluate the objective function. A typical objective function is maximization of efficiency for a given design point. In this work, we suggest to use CFD-trained metamodels which evaluate the objective function several orders of magnitude faster than CFD itself. The metamodels used are artificial neural networks (ANN) and differ from previously developed metamodels in terms of universality since they can be used for optimizing all typical design points of centrifugal impellers according to the Cordier diagram. In addition, the optimization scheme is supposed to handle typical operational and constructive constraints.
The biggest challenge in the development of the metamodels is to ensure adequate accuracy with limited computational resources required to generate the CFD dataset used to train the metamodels. For that purpose, the geometry was parameterized such that a large geometrical diversity (and hence the realization of very different design points) can be achieved with only nine independent geometrical parameters. The number of required geometry variations was reduced by a two-stage Design of Experiment consisting of a passive and an active learning phase. In the passive phase, the geometrical parameters were varied in a space-filling manner, i.e. only the inputs of the metamodels were considered. In the subsequent active phase, the aerodynamic performance (i.e. the metamodel outputs) was considered, too. The main target of the active phase was to focus mostly on those areas of the input space in which the (preliminary) metamodels have the lowest quality and on areas of high impeller efficiency since these areas are most relevant for the purpose of aerodynamic optimization. The computational time for the CFD simulations was kept to a minimum by using the comparatively cheap RANS method and by optimizing the computational grid with the aim of matching experimental performance data of three prototypes with as few grid points as possible.
The new optimization scheme was applied to numerous design points and the resulting geometries were simulated by CFD. It was found that the metamodels overpredict efficiency at design points with untypically small pressure coefficients. At standard design points and untypically large pressure coefficients, however, CFD confirms well the metamodel predictions proving the broad applicability of the new optimization scheme. In addition, four optimal geometries were manufactured and examined experimentally to validate the CFD method and to prove that the optimization scheme does not exploit weak points of the CFD model.