Using GPU-accelerated wall-modelled LES and ANN-based surrogate models for optimisation of installation effects in Distributed Propulsion Systems

Distributed Propulsion Systems (DPS) offer aerodynamic and noise benefits for Urban Air Mobility (UAM) vehicles, but their optimisation remains a challenge. This study employs GPU-accelerated Wall-Modelled Large Eddy Simulation (WMLES) and Artificial Neural Network (ANN)-based surrogate models to optimise DPS installation effects, focusing on two configurations: a single-propeller setup (A2) and a three-propeller configuration (B1). For A2, WMLES simulations conducted using the CABARET method examined propeller placement and inflow velocity effects, with a genetic algorithm optimising the setup to minimise Overall Sound Pressure Level (OASPL). The optimal configuration positioned the propeller at 777 mm with an inflow velocity of 26 m/s. In B1, WMLES simulated the influence of axial rotor positioning and phase angles on noise for both clean and serrated trailing-edge (TE) propellers. The ANN-based surrogate models predicted OASPL metrics, and multi-objective optimisation using NSGA-II identified optimal conditions for noise reduction. TE serrations were most effective when the central rotor was placed forward, with a 97° phase shift achieving the greatest OASPL reduction and a 5° shift minimising area-averaged OASPL.