Use of Physics Informed Neural Networks (PINNS) for Solving Fluid Flows

Abstract

Data obtained from experimental measurements often contain a large amount of noise and/or are of low resolution, so that the subsequent calibration of computational techniques can be complex. With the advent of machine learning techniques, deep learning approaches have been shown to be potentially suitable for improving images resolution, but their effectiveness is often limited by the need for large volumes of high-resolution reference data. In addition, the predictions generated by these networks may lack the necessary physical consistency, violating fundamental principles such as conservation of mass and momentum. In this sense, the incorporation of Physics Informed Neural Networks (PINNs) into the typical tasks scheme of Computational Fluid Dynamics (CFD) seems to be useful. Since PINNs allow physical laws (the governing equations) to be incorporated into the available data, it is possible to learn solutions that are intrinsically consistent with the physics of the problem being analyzed. In this work, PINNs are used to improve the resolution of fluid-flow data, using a limited set of noisy measurements and without the incorporation of high-resolution data as a reference.