Machine learning approaches for preliminary design of rotorcraft aeroacoustics

Abstract

Rotorcraft noise prediction remains challenging due to the large cost of experimental campaigns and high-order aeroacoustic simulations. This work develops a reduced-order, deep learning-based framework for the rapid synthesis of helicopter main-rotor noise hemispheres for preliminary design. A rotorcraft aero-mechanics and acoustics computational framework was used to generate noise hemispheres based on Latin Hypercube Sampling (LHS) across a wide range of design parameters and operating conditions. Two reduced-order modelling (ROM) methods were investigated: (i) POD+Kriging, combining dimensionality-reduction with Gaussian-process regression, and (ii) an enhanced U-Net Convolutional Neural Network (U-CNN) that reconstructs the full acoustic field directly from rotor design parameters. The ROMs were validated using cross-validation and evaluated on low-speed descent cases, a condition typically dominated by complex Blade-Vortex Interaction (BVI) noise. The POD-Kriging model achieves a mean root-mean-square error of 0.70 dB and the U-CNN achieves 0.53 dB across the unseen test configurations, recovering thickness and loading noise components even in BVI-dominated cases while maintaining the original resolution as a traditional computation method. Overall, the study establishes a novel framework for the rapid rotor noise characterisation suitable for rotorcraft preliminary design.