Multi-Fidelity Approaches for Predicting Propeller–Wing Interaction Noise

This study presents the development and evaluation of two multi-fidelity surrogate models for predicting the first blade-passage frequency acoustic directivity of a propeller-wing configuration across a parametric sweep of wing leading-edge positions. The configuration follows the experimental setup at NASA Langley Research Center, comprising a three-bladed Mejzlik propeller operating upstream of a NACA 632-215 MOD B wing at 40 discrete leading-edge positions spanning the horizontal and vertical parameter space. Medium-fidelity predictions from the Rotorcraft Comprehensive Analysis System (RCAS), using the Viscous Vortex Particle Method, serve as the low-fidelity input, while high-fidelity predictions from NASA's OVERFLOW solver coupled with PSU-WOPWOP, both serve as the training datasets. QR decomposition is employed in both frameworks to identify the most informative subset of wing positions for high-fidelity simulation. The first surrogate model, a linear regression model, requires 15 high-fidelity snapshots to outperform the RCAS predictions, achieving l∞ norm, l2 norm, and RMSEs below the RCAS baselines of 32.98%, 8.56%, and 5.67 dB, respectively, over the entire parameter space. The second surrogate model, a Co-Kriging surrogate model implemented via the Surrogate Modeling Toolbox 2.0, requires only seven high-fidelity snapshots and achieves an RMSE of 4.45 dB and an R² value of 0.67. Both surrogate models accurately reproduce the high-fidelity directivity at wing positions where the CFD and RCAS trends are in reasonable agreement, but struggle at positions where the two solvers predict substantially different trends, highlighting the sensitivity of multi-fidelity methods to solver discrepancies and the potential need for denser parameter space sampling in regions of high acoustic variability.