Multi-Fidelity Design of Multi-Curvature Fan Blade Based on Machine Learning of Vortex Strength and Efficiency

초록

In the research we develop and implement a design software based on multi-fidelity strategy for generating fan blade geometry in the design space to achieve optimal design configurations. Our approach is based on combining M.L. high-fidelity CFD simulations with a large number of lowfidelity computationally inexpensive models, to achieve minimal non-dimensional overall vorticity level and maximum efficiency near an operation point. While the high-fidelity CFD can provide detailed physics of unsteady vortex flows and aero-acoustic noise generation sources in detail, it is too computationally demanding for performing optimization or uncertainty quantification (UQ) studies over the range of parameters. This is because performing optimization and/or UQ requires a large number (often in the order of 100-10,000) of simulations for the system optimization for different configurations. In these cases, lower-fidelity models, with 1.5D mean-line design approximations, are used instead. However, these lower fidelity models approximate the underlying physics, and that leads to surface geometries with the quantities of interest within some degree of inaccuracies, but serving as a base domain of geometric parameters in order to activate high-fidelity simulation. It is therefore useful to develop multi-fidelity design software that combines the desirable characteristics of both high- and low-fidelity approaches. That is, they inherit the computational efficiency of the low-fidelity model, while retaining the accuracy of the high-fidelity model. The low fidelity approach involves morphing of various surfaces based on different vorticity distribution functions. In this paper the software performance of multifidelity design is tested and verified for generating high-efficient, low-noise automobile axial cooling fan blade geometry.