Distortion Correction In Particle Image Velocimetry For Measurements Through Fluctuating Interfaces Using A Deep Neural Network And Wavefront Sensing

Abstract

Imaging-based flow measurements like Particle Image or Particle Tracking Velocimetry re-quire an undistorted optical access to the measurement location. Changes of the refractive index in the light path can cause image deterioration and lead to severe measurement errors. While static aberration as e.g. from curved walls can be corrected by of means static correction optics or dedicated light ray models, this is hardly possible if the aberration change with time at high bandwidth. Many cases of this kind exists in fluid mechanics as e.g. free liquid jets, bubbles, droplets or film flows. These interfaces are related to several open research questions that concern the interaction between the media on both sides of the interface. Hardware-based corrections adapted from astronomy that employs a deformable membrane mirror, a wavefront sensor and a fast control electronics are a viable approach for solving this measurement problem. However, these closed-loop control schemes require extensive and dedicated hardware and are limited by the technical specifications of the adaptive optical mirror in terms of correction bandwidth, stroke and spatial resolution. Here we present an approach based on Multiple-Input deep convolutional neural network that works with a wavefront sensor only and spares the adaptive optical component. Based on a model experiment and a PIV measurement, we demonstrate that this system is able to correct for the errors that are induced by an oscillating water surface by 82 %. The technique can be an alternative to classical closed-loop adaptive optical systems where the performance of the actuators is not sufficient. This can open a perspective towards measurement of interface-related flows that were not accessible before.