Influence of adversarial training on super-resolution turbulence models

Abstract

Supervised super-resolution deep convolutional neural networks (CNNs) have gained significant attention for large eddy simulation subfilter-scale (SFS) modeling due to their ability to reconstruct statistically meaningful flow fields on fine meshes. Despite their popularity, CNNs lack the ability to accurately reconstruct high-frequency features and generalization performance on out-of-sample flows. Generative adversarial networks (GANs) are a potential alternative, allowing for both semi-supervised and fully unsupervised training, though they have not been thoroughly investigated as turbulence closures, and a comprehensive understanding of the discriminator's role has not been developed. This study assesses the effectiveness of GANs for a priori SFS stress modeling in forced homogeneous isotropic turbulence. It is found that GAN-based architectures outperform supervised CNN models for SFS reconstruction for in-sample cases. The reconstruction accuracy of both models decreases for out-of-sample data, though the GAN discriminator applied as a "feature extractor" narrows the model's solution space and enhances the generator's out-of-sample robustness. The extrapolation ability of the GAN-based model for higher-Reynolds-number flows is also demonstrated. This highlights the effectiveness of the GAN discriminator in optimizing robust and accurate SFS models for out-of-sample flows. Based on these findings, training with a discriminator is recommended before integrating super-resolution CNN closures into numerical solvers.

More information

Identifiers

ORCID for T. Grenga: orcid.org/0000-0002-9465-9505

Catalogue record

Export record

Altmetrics