Multifidelity surrogate presented at CEAS Torino

The first work package of ENOLA is focused on multi-fidelity propeller modelling, testing and optimization. In this recent work, presented at the AIDAA-CEAS conference in Torino, multi-fidelity numerical acoustic optimization of UAV propellers, we present a unified methodology to efficiently optimize fixed-pitch propellers operating at low Reynolds numbers. The proposed approach integrates Blade Element Momentum Theory (BEMT) coupled with the Hanson formulation for tonal noise as a low-fidelity model, and a Vortex Particle Method (VPM) coupled with an FW–H acoustic solver as a higher-fidelity alternative. The model is also prepared for high-fidelity VPM and FVM (CFD) simulations. All are driven by a common geometric definition, enabling automated multi-level simulations from the same propeller description. The methodology was experimentally validated in the anechoic chamber at CMT using a 9-inch commercial propeller, showing good agreement in blade-passing-frequency directivity patterns and highlighting the benefits of higher-fidelity simulations for upstream noise prediction.

To build the surrogate, or Reduced Order Model (ROM), we trained a compact fully connected neural network surrogate that maps propeller geometry and thrust requirements to rotational speed, torque, and sound pressure level at multiple observer positions. Three datasets were considered: BEMT-only, VPM-only, and a balanced mixed set. While the BEMT surrogate achieved very low in-domain errors, the mixed-fidelity model provided a more robust representation across simulation schemes, maintaining sub-1.5 dB errors at most observer locations while generalizing better to higher-fidelity trends.

Finally, the surrogate was embedded in a genetic optimization framework to generate Pareto fronts balancing torque and tonal noise at blade passing frequency. The results show meaningful differences between single- and multi-fidelity optimization outcomes, underlining the importance of accounting for fidelity-dependent effects during the design phase. Ongoing work focuses on reducing residual discrepancies—particularly in torque prediction—and on incorporating transfer learning strategies to further bridge the gap between low- and high-fidelity models, strengthening the digital toolchain for reducing UAV noise, the final objective of ENOLA.