Application of computer vision and neural network analysis to study the structure and dynamics of turbulent jets

Abstract

In this paper we solved the problem of detecting and detailed analysis of the characteristics of coherent structures in non-swirling and swirling turbulent jets from the 3D PIV experiment and DNS modelling database. To solve the problem intellectual approaches of computer vision based on deep neural networks were applied. It is shown that with the use of a generative competitive neural network, it is possible to reconstruct the structure and dynamics of a turbulent flow with an increased spatial resolution, which is important for analysis and interpretation. The approach is based on a modification of the loss function, which minimizes the residual part of the hydrodynamic equations (the continuity equations and the Poisson equation) for the reconstructed data. It is shown that generative models are more effective in the reconstruction of high moments of turbulent flow than conventional methods POD and DMD. Using a fully convolutional neural network, an automatic segmentation of the turbulent jet flow region and the external environment based on instantaneous velocity and pressure fields was carried out. For a swirling jet, the area of reversed flow is identified. It is shown that the complex of the developed machine learning algorithms successfully copes with the localization of coherent structures, the detection of their trajectory, characteristic size and phase propagation velocity. The obtained new fundamental information is important for a deeper understanding of the role of vortex structures in mixing processes in a jet flow. The developed complex of algorithms for the identification and analysis of characteristics of coherent structures can be applied to a database of measurements and numerical simulation of a wide class of hydrodynamic flows.