Deep Learning Surrogate Model for Rotorcraft Aeroacoustic Simulations using a Geodesic Convolutional Network

Abstract

This paper presents a new state of the art surrogate assisted aeroacoustics toolchain, with a deep learning surrogate model able to accurately predict blade pressure distributions for helicopter main rotors in various flight conditions. By training a state of the art geodesic convolutional neural network, the surrogate model is able to accurately predict new cases in 0.1s with an R2 of 99% - 10,000 faster than the existing tool. Surrogate models have been built for four LH platforms, and a multi-blade model developed which showed very promising rotor performance predictions for an unseen blade design. The helicopter aeroacoustic tool chain can now run in just 8 minutes, down from 8 hours. In future work, the training of the surrogate models will be extended, to consider a wider flight envelope of flight conditions, and grow its ability to predict unseen rotor blade performance.