Abstract
Three methods are used to reconstruct two-dimensional instantaneous velocity fields in a turbulent flow under rotation. The first two methods both use the linear proper orthogonal decomposition (POD), which are Gappy POD (GPOD) and Extended POD (EPOD), while the third one reconstructs the flow using a fully non-linear Convolutional Neural Network embedded in a Generative Adversarial Network (GAN). First, we show that there is always an optimal number of modes regarding a specific gap for the GPOD with dimension reduction. Moreover, adopting a Lasso regularizer for GPOD provides comparable reconstruction results. In order to systematically compare the applicability of the three tools, we consider a square gap at changing the size. Results show that compared with POD-based methods, GAN reconstruction not only has a smaller $L_2$ error, but also better turbulent statistics of both the velocity module and the velocity module gradient. This can be attributed to the ability of nonlinearity expression of the network and the presence of adversarial loss during the GAN training. We also investigate effects of the adversarial ratio, which controls the compromising between the $L_2$ error and the statistical properties. Finally, we assess the reconstruction on random gappiness. All methods perform well for small- and medium-size gaps, while GAN works better when the gappiness is large.
Abstract (translated)
URL
https://arxiv.org/abs/2210.11921
