Abstract
The behaviours of individual flow structures have become a relevant matter of study in turbulent flows as the computational power to allow their study feasible has become available. Especially, high instantaneous Reynolds Stress events have been found to dominate the behaviour of the logarithmic layer. In this work, we present a viability study where two machine learning solutions are proposed to reduce the computational cost of tracking such structures in large domains. The first one is a Multi-Layer Perceptron. The second one uses Long Short-Term Memory (LSTM). Both of the methods are developed with the objective of taking the the structures’ geometrical features as inputs from which to predict the structures’ geometrical features in future time steps. Some of the tested Multi-Layer Perceptron architectures proved to perform better and achieve higher accuracy than the LSTM architectures tested, providing lower errors on the predictions and achieving higher accuracy in relating the structures in the consecutive time steps.
Highlights
In this paper, the terms “turbulent structures”, “Qs” and “sweeps, and ejections” are used interchangeably.Turbulence is probably the open problem in physics with most applications in daily life
The work of Srinivasan et al in predicting boundary layer behaviour by means of convolutional and recurrent neural networks [18], the work of Park and Choi in Energies 2021, 14, 984 reconstructing data for flow control by means of convolutional neural networks [19], the work of Ling et al in using Artificial Neural Networks (ANNs) to estimate the Reynolds Stress anisotropy tensor to augment the results provided by Reynolds Average Navier Stokes (RANS) models [20], and other ANN approaches to modelling the closure problem [21,22] are good examples of this
We have compared the performance of several Multi-Layer Perceptrons (MLP) and Long Short-Term Memory (LSTM) artificial neural networks in predicting the future geometrical features of the sweeps and ejections of a turbulent channel flow at a relatively low friction Reynolds number, Reτ = 500
Summary
Turbulence is probably the open problem in physics with most applications in daily life. To completely understand turbulence is perhaps a too-ambitious problem, which we still are years apart, so we should focus on improving models such that they can simulate, in a reliable and fast way, the behaviour of flows. Because we want to understand these flows first, no modelling can be used, so we have to used the Direct Numerical Simulation (DNS) technique. This technique does not use any other modelling than the Navier–Stokes equations. Since the 90’s of the last century, their power have increased in an exponential way, and the DNS have grown . Because the seminal work of Kim, Moin, and Moser [2], the main control parameter, the friction Reynolds number Reτ has increased continuously [2,3,4,5,6,7,8]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
