Development, Application, and Validation of a Quick Optimization Method for the Class of Axial Fans

Abstract

This article discusses the development, application, and validation of an optimization method for the impellers of axial fans. The method is supposed to be quick, accurate, and applicable to optimization at an extensive range of design points (DPs). Optimality here means highest possible total-to-static efficiency for a given design point and is obtained by an evolutionary algorithm in which the target function is evaluated by computational fluid dynamics (CFD)-trained artificial neural networks (ANN) of the multilayer perceptron (MLP) type. The MLPs were trained with steady-state CFD (i.e., Reynolds-averaged Navier–Stokes (RANS)) results of approximately 14,000 distinct impellers. After this considerable one-time effort to generate the CFD dataset, each new fan optimization can be performed within a few minutes. It is shown in this article that the MLPs are reliably applicable to all typical design points of axial fans according to Cordier's diagram. Moreover, an extension of the design space toward the classic realm of mixed-flow or even centrifugal fans is observed. It is also shown that the optimization method successfully handles geometrical and operational constraints proving the high degree of universality of the method. Another focus of this article is on the application of the newly developed optimization method to numerous design points. This yields two major findings: the estimation of maximum achievable total-to-static efficiency as a function of the targeted design point (with and without geometrical constraints) as well as a quantification of the improvement over fans designed with classic methods. Both investigations are supported by flow field analyses to aerodynamically explain the findings. Experimental validation of the method was performed with a total of nine prototypes. The positive correlation between MLP, CFD, and experiment successfully validates the methodology.