De-noising and super-resolution of experimental velocity measurements by optimizing a discrete loss (ODIL)

Abstract: This article presents a comparison between two techniques optimising a discrete loss (ODIL) and physics informed neural networks (PINN) for reconstructing the velocity field from low resolution and noisy planar PIV measurements. Both techniques are capable of accurately reconstructing the velocity and pressure fields from low resolution and noisy velocity measurements sampled from 2D numerical simulations. Both techniques provide reasonable reconstruction of the in-plane velocity components when provided with two-component velocity measurements in a single plane sampled from a 3D numerical simulation. However, ODIL generally over-fits to any noise in the measurement data, while PINN may under-fit the data, and not reconstruct small-scale flow structures. In all cases, ODIL requires fewer training epochs and less wall-time than PINN to produce results of similar accuracy. Keywords: machine learning, data assimilation, flow reconstruction