Multi-Fidelity Artificial Neural Networks For Rotorcraft Airfoil Design

Abstract

Rotor blade design is a fundamental problem in the field of rotorcraft aeromechanics. The conventional design methods using high-fidelity techniques (like Computational Fluid Dynamics (CFD)) are computationally expensive and often become impractical at the early design phase. The use of deep learning techniques for blade design is an efficient alternative to conventional design methods. However, these machine learning techniques often require a lot of training data which are generated using the CFD. The training data generation process becomes a bottleneck and requires a lot of computational effort, before being able to train the neural networks. In this paper, we propose and implement a unique multi-fidelity artificial neural network (MFANNs) architecture that takes low-fidelity (Xfoil) data as inputs and maps/transforms it to high-fidelity data, requiring 1/10 of high-fidelity training data as compared to the conventional high-fidelity artificial neural networks. A comprehensive study on the number of high-fidelity cases required for training MFANNs and the sensitivity analysis of data reduction is also presented in the paper. The multi-fidelity architecture is integrated with an evolutionary algorithm to optimize the airfoil.