Super-resolution of Turbulent Flows in the Absence of the High-Resolution Ground Truth

Abstract

Super-resolution (SR) refers to the problem of reconstructing high-resolution (HR) information from low-resolution (LR) data. SR of physical data has attracted a great attention with the advent of deep learning. While most of the SR models assume existence of HR ground truth data to train a deep learning model, in real-life problems, such HR ground truth data is scarce. Here, we present a deep learning model for SR of turbulent flows trained without such HR ground truth. We propose to use the conservation laws to regularize the solution of the SR model. It is shown that the SR model can reliably reconstruct unseen high- resolution turbulent flows.