Performance Prediction of Axial Fans by CFD-Trained Meta-Models

The work presented in this paper is concerned with a methodology for substituting time consuming CFD investigations of the operational characteristics of axial fans by CFD-trained meta-models. For that, the fan geometry is parameterized by 25 physically interpretable quantities allowing for a huge variety of potential fan designs. The parameters are varied by Design of Experiment (DoE) and characteristic curves of approximately 10,000 fan designs are produced using the Reynolds-averaged Navier Stokes (RANS) method. Pressure rise, efficiency, and circumferentially averaged flow profiles upstream and downstream of the rotor are extracted from the RANS results and used to train the meta-models which are Artificial Neural Networks (ANN) or, more specifically, Multilayer Perceptrons (MLP). Special care is taken to mitigate extrapolation weaknesses of the MLPs which could compromise their suitability to compute the target function in optimization algorithms. With these extra efforts, it is possible to aerodynamically optimize axial fans for arbitrary design points within the range typical for axial or even mixed-flow fans according to Cordier’s diagram of turbo machinery. On top of that, designs with good efficiency are also found outside the well known Cordier range. In particular, an extension of feasible operating points towards untypically high specific fan diameters is observed. These findings are relevant for designs aiming at high total-to-static efficiency and make optimized axial fans compete with other fan types, especially with mixed-flow fans.