Blade Shape Optimization of Rotors using Neural Networks

This paper presents a methodology for using a neural network to predict airfoil behavior in the context of rapid airfoil design and prop-rotor blade shape optimization. To train the neural network, the 1620 shapes in the UIUC airfoil database were each evaluated using XFOIL at various angles of attack. The Class Shape Transformation approach is used to parameterize airfoil upper and lower surface geometries, using Chebyshev polynomials as shape functions. Three separate neural networks were trained, one each for lift, drag, and pitching moment. After training the neural network, isolated airfoil shape optimization was performed using the NSGA2 algorithm, targeting minimum average drag over an operating lift coefficient range for 10%, 12%, and 16% thick airfoils. Additionally, prop-rotor aerodynamic optimization was carried out by designating airfoil shape parameters, blade twist distribution, and blade chord distribution as simultaneous design variables for two objectives: hover figure of merit and cruise-mode propeller efficiency. A Blade Element Momentum Theory is used to predict rotor performance, using airfoil tables generated by the neural network. Pareto frontiers and an analysis of the resulting designs are presented. Using a neural network is advantageous for both applications considered, because it results in a 40 × speed-up over XFOIL, and decouples the computational cost of shape optimization from that of the physics-based model generating truth data.