Influence of adversarial training on super-resolution turbulence reconstruction

Abstract

We compare supervised super-resolution convolutional neural networks (CNNs) against generative adversarial networks (GANs)-based architectures in the ability to reconstruct turbulent flow fields. GANs demonstrated superior in-sample performance but faced challenges with out-of-sample flows. Incorporating a partially unsupervised adversarial training step with large eddy simulation inputs and dynamic upsampling selection improved GANs' out-of-sample robustness, capturing small-scale features and turbulence statistics better than standard supervised CNNs. The study recommends integrating discriminator-based training to enhance super-resolution CNNs' reconstruction capabilities.