Physics-Informed Machine Learning and 2-Component Schlieren and Shadowgraph Imaging of a 2-D Soap Film

Abstract

In this paper, a contactless combined optical and computational approach to inferring flow variables in a dynamic 2-D soap film is demonstrated. Using a novel method for simultaneous high-speed two-component schlieren and shadowgraph imaging, dynamic film thickness gradient profiles for vortex shedding and comb turbulence are acquired from a soap film generated in an experimental soap film tunnel. For vortex shedding, we then demonstrate significant denoising of our data using an autoencoder and simultaneous boundary-condition-dependent integration and experimental error correction using a physics-informed neural network. Modeling film thickness as an advected scalar quantity and considering the joint system of PDEs: Navier-Stokes and advection-diffusion, we then apply physics-informed machine learning to learn the simultaneous dynamic velocity and pressure fields. While this work is ongoing, we present encouraging results for velocity field predictions. Additional soap film thickness measurement techniques: thin-film interference and $3\ \mu m$ IR absorption imaging, are discussed as well.