In many areas of data science, deep neural networks (DNNs) have shown a remarkable ability to learn complex, nonlinear relationships between sets of variables. In this paper, this network architecture is applied to several different tasks relating to high-speed turbulent flows. In the first section, linear stochastic estimation (LSE) as proposed by Adrian (“On the Role of Conditional Averages in Turbulence Theory,” Symposium on Turbulence in Liquids, 1977; and “Conditional Eddies in Isotropic Turbulence,” Physics of Fluids, Vol. 22, No. 11, 1979, pp. 2065–2070) is reformulated as a machine learning problem, and the two methods are compared. Both a DNN and a LSE model are trained to estimate fluctuating pressure at a subset of locations in the near field of a Mach 0.6 jet, given the pressure measured at other locations. It is shown that DNNs exhibit a slight performance benefit over traditional LSE models on average. The second part of this paper focuses on the utilization of an artificial neural network (ANN) to predict the directional overall sound pressure level (OASPL) in the far field of a supersonic multistream jet. A database was created, describing the near-field and far-field conditions of a complex nonaxisymmetric jet flow, with Mach numbers ranging from 1.0 to 1.6. The problem was posed as a form of multivariate nonlinear regression, and an ANN was used to create a model. A feature space consisting of plausible predictors of the far-field directional OASPL was defined, based on previous fundamental studies and jet noise scaling laws. On average, the ANN was able to predict the directional far-field OASPL within 0.75 dB, surpassing original goals. In addition to these topics, some limitations and possible extensions of the methods described herein are discussed.
ASJC Scopus subject areas
- Aerospace Engineering