Neural network noise prediction for a complex supersonic rectangular jet nozzle

A method for the prediction of far-field acoustics emanating from a complex supersonic rectangular nozzle using machine learning techniques is presented. Complexity and lengthy times associated with traditional experimental and numerical procedures motivate the use of a rapid noise prediction system based on modern machine learning computational methods. An artificial neural network (ANN) is employed to predict the far-field overall sound pressure levels (OASPL) for a Multi Aperture Rectangular Single Expansion Ramp Nozzle (MARS). A combination of experimental techniques and machine learning algorithms are utilized. The results yield a simple, accurate, and computationally frugal method for the prediction of far-field acoustic magnitudes and directions. Realistic operating conditions and geometric properties serve as inputs to this model which calculates OASPL at multiple far-field locations. This provides a basis for further exploration of geometric modifications and passive control schemes while continuing to explore the effects of a newly installed splitter plate trailing edge (SPTE) geometry. Preliminary Large Eddy Simulation (LES) data performed by The Ohio State University have shown that the introduction of a spanwise wavenumber to the SPTE results in a reduction of the 34 kHz tone associated with the shedding frequency of the splitter plate. With this dominating tone greatly reduced, a more in depth study into the acoustic effects of different aft deck lengths can be performed and ultimately optimized for a desired acoustic output.