Experimental characterisation and data-driven modelling of unsteady wall pressure fields induced by a supersonic jet over a tangential flat plate

Abstract

This work deals with the investigation and modelling of wall pressure fluctuations induced by a supersonic jet over a tangential flat plate. The analysis is performed at several nozzle pressure ratios around the nozzle design Mach number, including slightly over-expanded and under-expanded conditions, and for different radial positions of the rigid plate. Pitot measurements and flow visualizations through the background oriented schlieren technique provided a general overview of the aerodynamic interactions between the jet flow and the plate at the different regimes and configurations. Wall pressure fluctuations were measured using a couple of piezoelectric pressure transducers flush mounted over the plate surface. The spectral analysis has been carried out to clarify the effect of the plate position on the single and multivariate wall pressure statistics, including the screech tone amplitude. The experimental dataset is used to assess and validate a surrogate model based on artificial neural networks. Sound pressure levels and coherence functions are modelled by means of a single fully connected network, built on the basis of a recently implemented fully deterministic topology optimization algorithm. The metamodel uncertainty is also quantified using the spatial correlation function. It is shown that the flow behaviour as well as the screech and broadband noise signatures are significantly influenced by the presence of the plate, and the effects on spectral quantities are correctly reproduced by the proposed data-driven model that provides predictions in agreement with the available data.

An Extensive Near-Field Noise Prediction of a Subsonic Jet Using Data-Driven Surrogate Model Based on Neural Networks

“…2. The selection of weight matrices {W 𝑙 } 𝐿-1 𝑙=1 and bias vectors {b 𝑙 } 𝐿 𝑙=2 is performed by means of the so-called learning, and is here addressed through the so-called training with back-propagation method: a cost function is computed at the 𝑡-th epoch (the iteration of the training process), so that its gradient is used to update weight and bias values, back-propagating the error information layer by layer [14]. Weights and biases at the 𝑡-th epoch are updated used the learning rate, a free parameter selected by the designer to control the change in weights and biases based on the error gradient components.…”

“…Centracchio et al [13] recently proposed a data-driven nonlinear model based on ANNs to describe and predict the noise emitted by a single stream jet in under-expanded conditions. A metamodel on the wall pressure fluctuations generated by an installed supersonic jet has been carried out by Meloni et al [14], Iemma et al [15], while the prediction of the noise directivity generated by an ingesting boundary layer propeller has been presented by Meloni et al [16]. Machine learning has been successfully exploited in similar configurations, for example, to correlate computational fluid dynamics data to the jet acoustic response [17] or to successfully predict the far-field noise spectra produced by supersonic jets with different nozzle shapes located near a surface [18].…”

“…Following the approach presented in Meloni et al [14], we aim at providing a prediction of the sound pressure level with a low uncertainty in a large numerical domain which spans from the nozzle exhaust up to 𝑥/𝐷 = 20 in the streamwise direction and up to 𝑟/𝐷 = 3 in the radial direction. To tune the present model, we used a numerical dataset obtained from the Large-Eddy Simulation (LES) of an isothermal round free jet at a Mach number of 0.9 and a diameter-based Reynolds number of 𝑅𝑒 𝐷 = 10 5 varying the nozzle-exhaust turbulence level between 𝑇 𝐼 = 0% (fully laminar flow) and 𝑇 𝐼 = 15%, which is representative of a fully turbulent jet, and the boundary layer thickness at the nozzle exit between 𝛿 𝐵𝐿 = 0.05 𝑟 0 and 𝛿 𝐵𝐿 = 0.4 𝑟 0 , where 𝑟 0 is the jet radius.…”

Physics-agnostic and physics-infused machine learning for thin films flows: modelling, and predictions from small data

“…Also, data-driven modelling has recently been proposed in the context of unsteady wall pressure fields (Meloni et al. 2023).…”

“…A model based on a convolutional neural network is proposed so as to reconstruct the three-dimensional turbulent flows beneath a free surface using surface measurements, including the surface elevation and surface velocity (Xuan & Shen 2023). Also, data-driven modelling has recently been proposed in the context of unsteady wall pressure fields (Meloni et al 2023).…”