Deep Learning for Viscoelastic Fluids and Turbulent Diffusion

Abstract

This study aims to predict the conformation stress field of viscoelastic, drag-reducing turbulent channel flow and to estimate the passive-scalar diffusion in Newtonian turbulent channel flow. In general, understanding of the nature of wall turbulence and predicting of turbulent phenomena have been based on statistical features, but here we would discuss the possibility of extracting non-trivial features from instantaneous fields with the help of machine learning. If we can extract features unique to the instantaneous field, which are lost when equalized by statistical processing, it could be possible to predict the interior (intrinsic physics) and exterior (initial and/or boundary conditions) from limited spatial, temporal information. Using deep learning of CNN such as U-Net, we demonstrated the reproduction of the conformation stress field from the instantaneous velocity field and the estimation of the diffusion point source location from the downstream concentration distribution based on instantaneous local information. We have successfully demonstrated the usefulness of instantaneous local information, which might be lost in statistics, and confirmed the possibility of using deep learning to construct a viscoelastic-fluid surrogate model and a diffusion-source estimation tool.