A Data-Driven Reduced-Order Model for Installed Propeller Noise Prediction

Abstract

This research presents a data-driven approach to efficiently predict tonal noise generated by wing-installed propellers. We developed a workflow integrating first-principles aerodynamic simulations (UPM) with the Ffowcs Williams-Hawkings (FWH) equation-based solver (APSIM), used to generate a large-scale dataset for training a reduced-order model. Two vortex-based aerodynamic methods, the vortex filament method (VFM) and the vortex particle method (VPM), were evaluated; VPM demonstrated superior accuracy for installed configurations and was selected for data generation. Sensitivity studies identified ten key design and operating parameters influencing far-field noise. UPM-APSIM simulations were performed across a Halton-sequenced design space to create a comprehensive dataset. This data was then used to train a fully connected neural network (FCNN), serving as our reduced-order model (ROM). The trained ROM was validated against fly-over measurements from a DLR Dornier DO-228 aircraft, demonstrating good agreement in predicting tonal noise levels for the first five harmonics. This data-driven approach offers a computationally efficient means of predicting propeller noise, significantly faster than traditional methods. Future work will focus on expanding the dataset and incorporating higher-fidelity data to improve the model’s predictive capabilities across a broader range of operating conditions and frequencies.