Velocity field reconstruction of mixing flow in T-junctions based on particle image database using deep generative models

Abstract

Flow field data obtained by particle image velocimetry (PIV) could include isolated large damaged areas that are caused by the refractive index, light transmittance, and tracking capability of particles. The traditional deep learning reconstruction methods of PIV fluid data are all based on the velocity field database, and these methods could not achieve satisfactory results for large flow field missing areas. We propose a new reconstruction method of fluid data using PIV particle images. Since PIV particle images are the source of PIV velocity field data, particle images include more complete underlying information than velocity field data. We study the application of PIV experimental particle database in the reconstruction of flow field data using deep generative networks (GAN). To verify the inpainting effect of velocity field using PIV particle images, we design two semantic inpainting methods based on two GAN models with PIV particle image database and PIV fluid velocity database, respectively. Then, the qualitative and quantitative inpainting results of two PIV databases are compared on different metrics. For the reconstruction of velocity field, the mean relative error of using the particle image database could achieve a 52% reduction compared to a velocity database. For the reconstruction of vorticity field, the maximal and mean relative errors can reduce by 50% when using the particle image database. The maximum inpainting errors of two database inputs are both mainly concentrated on the turbulence vortex area, which means the reconstruction of complex non-Gaussian distribution of turbulence vortex is a problem for semantic inpainting of the experimental data.