Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : November 07, 2019 - November 08, 2019
We carried out direct numerical simulation (DNS) of pulsating turbulent pipe flow and predicted the time evolution of the flow field by deep learning. Pulsation control is one of the turbulence control methods to realize drag reduction. A rapid prediction of pulsating flows by deep learning is beneficial to optimize the pulsation patterns. The deep learning model in the present study consists of a convolutional autoencoder (CAE) and two-layers of long short-term memory (LSTM). The training data are images of velocities and pressure field in a cross-section which is calculated by the DNS. The feature vectors are extracted from the images by the CAE. The time-dependency of the flow dynamics is learned by the LSTM. The spatially averaged pressure gradient that is given to pulsate the flow is concatenated to the feature vector. By applying sequence-to-sequence learning to the model to predict the long-term dynamics, the model successfully reproduced the time evolution of the distribution of velocity and pressure fields and the flow statistics. The temporal variation of the friction coefficient calculated from the predicted flow field is approximately identical to those of the DNS. The relative error of the time-averaged friction coefficient is 6.0%. The model roughly predicted the friction drag of the pulsating flow.
