Predictability study of viscoelastic turbulent flow with deep learning

Abstract

We tested Artificial Neural Networks (ANNs) to predict a fully-developed turbulent channel flow of a viscoelastic fluid in preparation for elucidating flow phenomenon and solving the difficulty in DNS (Direct Numerical Simulation) due to numerical instability of the viscoelastic fluid. Two kinds of ANNs (multi-layer perceptron (MLP) and U-Net) were trained using DNS data to predict conformation stresses from given instantaneous fields. The MLP predicted the same tendency with DNS results and provided a better prediction with z-score normalization. The U-Net failed in accurate prediction even quantitatively, but both MLP and U-Net still suffered from the overfitting to the time slot of training data.