From Sparse to Dense Representations in Open Channel Flow Images with Convolutional Neural Networks

Abstract

Convolutional neural networks (CNN) have been widely adopted in fluid dynamics investigations over the past few years due to their ability to extract and process fluid flow field characteristics. Both in sparse-grid simulations and sensor-based experimental data, the establishment of a dense flow field that embeds all spatial and temporal flow information is an open question, especially in the case of turbulent flows. In this paper, a deep learning (DL) method based on computational CNN layers is presented, focusing on reconstructing turbulent open channel flow fields of various resolutions. Starting from couples of images with low/high resolution, we train our DL model to efficiently reconstruct the velocity field of consecutive low-resolution data, which comes from a sparse-grid Direct Numerical Simulation (DNS), and focus on obtaining the accuracy of a respective dense-grid DNS. The reconstruction is assessed on the peak signal-to-noise ratio (PSNR), which is found to be high even in cases where the ground truth input is scaled down to 25 times.