A Multi-Fidelity, Multi-Start Approach For The Aerodynamic Design Of Propellers

Abstract

This paper presents a new approach for aerodynamic shape design and optimisation, combining high-fidelity gradient-based optimisation with multi-fidelity surrogate-based optimisation. It exploits the gradient-based optimisation history (both function values and gradients) to train gradient-enhanced multi-fidelity surrogate models, and uses the surrogate to indicate potentially better global solutions to restart the gradient-based local search. This multi-fidelity, multi-start (MFMS) approach retains the high efficiency of gradient-based methods in searching high-dimensional design spaces and helps evade suboptimal local solutions via surrogates. A key novelty here is the introduction of a multi-fidelity neural network (MFNN), which seamlessly fuses physical models, multi-fidelity data, and gradients, providing accurate surrogate predictions on very sparse samples. The proposed MFMS aerodynamic shape optimisation framework was verified and evaluated via benchmark supercritical aerofoil shape optimisation. The framework was further demonstrated using the HART II rotor optimisation case as part of the International Design Workshop on Rotor Blade Optimisation (InDeWo).