Experimental measurement-based frequency-conditioned neural network modeling of conical acoustic liners under combined grazing and bias flow

Abstract

Acoustic liners are essential for noise reduction in aeroengines, but geometric parameter optimization for these in complex flows still represents an important challenge. The work presents a machine learning approach to estimate and optimize sound absorption characteristics for conical perforated liners with a conical cavity in the presence of complex flows, including grazing and bias flows. The present work has compiled an experimental database for 81 cases with Mach numbers up to 0.1 using the ASTM E2611-19 two-load approach in an impedance tube, with a frequency span between 500 and 2500 Hz. These data have been used to train the deep feed-forward neural network, and comparisons between the results of different optimizers, including Adam, stochastic gradient descent (SGD) with Nesterov acceleration, and the recently proposed AdamW methods, are reported in detail based on root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) for the estimation of sound absorption characteristics. The results clearly highlight the superiority of the adaptive methods with respect to their robustness and convergence properties for reproducing the frequency dependencies of sound absorption in these complex flows, and the described approach appears very appealing for sound treatment in complex flows, providing a fast and powerful solution in comparison with present experimental and numerical approaches.