Generative adversarial networks for turbulence modeling in large-eddy simulation

RWTH Publications (RWTH Aachen)

Abstract

Generative adversarial networks (GANs), which jointly train a generator and a discriminator, have excelled in super-resolution (SR) applications, yet their suitability for large-eddy simulation (LES) modeling remains largely unexplored. This dissertation systematically evaluates SR-GAN–based closures for forced homogeneous isotropic turbulence with a dual focus on predictive accuracy and computational efficiency. A priori and a posteriori studies demonstrate that the enhanced performance of SR-GAN over conventional fully supervised SR approaches is attributable solely to adversarial training. As the discriminator learns to distinguish super-resolved from DNS fields, high-order DNS statistics are encoded in its latent space and act as an implicit, statistical, physics-informed loss function for the generator. This yields accurate in-sample reconstructions and improved out-of-sample robustness of the generator. To address the substantial computational demands of SR-GANs, investigations on distributed, patch-based training and hybrid CPU–GPU inference are performed. The results indicate that accurate reconstructions require patches sufficiently large relative to the Kolmogorov scale and careful balancing of generator–discriminator feedback. Similarly, a posteriori SR-coupled LES imposes restrictions on domain decomposition due to convolutional boundary effects, which limit parallel scalability and increase computational cost. As a result, employing SR-LES for full-scale (DNS-like) subfilter-scale reconstruction is impractical even for moderate Reynolds-number flows. To mitigate these computational limitations, a dynamic mixed super-resolution model (DMSRM), which employs the traditional dynamic mixed model (DMM) formulation on a super-resolved velocity field upsampled by a factor of two, is proposed. DMSRM consistently improves the energy spectrum, intermittency, and subfilter-stress magnitudes and dissipation over DMM, even across different LES grid-scale resolutions and for higher Reynolds numbers than those used for training. These improvements arise mainly from the adversarial training enabled by the GAN framework, which yields physically accurate super-resolved fields. Despite increased training difficulty and higher computational costs, the accuracy, robustness, and hybridization potential of SR-GANs make them a promising approach for improving LES closure models.

Authors 1

1. Ludovico Nista corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen