ZipGAN: Super-Resolution-based Generative Adversarial Network Framework for Data Compression of Direct Numerical Simulations

TL;DR

This paper tackles the challenge of storing and transferring massive DNS HIT datasets by introducing ZipGAN, a SR-GAN-based decoder that reconstructs high-fidelity 3D turbulence fields from compressed, downsampled inputs. The approach combines a SR-ResNet–style generator with RRDB blocks and a relativistic GAN discriminator, trained through a multi-term loss and progressive transfer learning to enable high compression ratios up to $\theta=512$ while preserving small-scale features and gradients. It outperforms traditional DWT baselines in reconstruction fidelity (NMSE, SSIM) and supports temporal upsampling, with a discriminator-based quality score $Q_s$ enabling assessment even for out-of-sample data. The method remains effective across Reynolds numbers when $dx/\eta$ is preserved, offering substantial storage savings (up to roughly 100x) and practical scalability for large-scale DNS datasets, aided by GPU-accelerated decoding and transfer-learning strategies.

Abstract

The advancement of high-performance computing has enabled the generation of large direct numerical simulation (DNS) datasets of turbulent flows, driving the need for efficient compression/decompression techniques that reduce storage demands while maintaining fidelity. Traditional methods, such as the discrete wavelet transform (DWT), cannot achieve compression ratios of 8 or higher for complex turbulent flows without introducing significant encoding/decoding errors. On the other hand, a super-resolution-based generative adversarial network (SR-GAN), called ZipGAN, can accurately reconstruct fine-scale features, preserving velocity gradients and structural details, even at a compression ratio of 512, thanks to the more efficient representation of the data in a compact latent space. Additional benefits are ascribed to adversarial training. The high GAN training time is significantly reduced with a progressive transfer learning approach and, once trained, they can be applied independently of the Reynolds number. It is demonstrated that ZipGAN can enhance dataset temporal resolution without additional simulation overhead by generating high-quality intermediate fields from compressed snapshots. The ZipGAN discriminator can reliably evaluate the quality of decoded fields, ensuring fidelity even in the absence of original DNS fields. Hence, ZipGAN compression/decompression method presents a highly efficient and scalable alternative for large-scale DNS storage and transfer, offering substantial advantages over the DWT methods in terms of compression efficiency, reconstruction fidelity, and temporal resolution enhancement.

Figures (6)

• Figure 1: The generator (decoder) and discriminator structures employed by the ZipGAN architecture. Input and output fields reported in Eq. \ref{['eqn:theory_compression_data']} are highlighted. Each convolutional block contains kernels of size $k$, $n$ filter maps, and $s$ strides along each spatial dimension of the convolutional layer. ZipCNN specifically refers to the generator of ZipGAN, which is trained in a fully supervised manner without the adversarial contribution present in the complete GAN framework.
• Figure 2: PDF of the normalized velocity gradient of the decoded field, $\mathrm{\phi_{dec}}$, obtained for different compression ratios, $\mathrm{\theta}$, using DWT (left) and ZipGAN (right) compression methods on the Re90 dataset. The input encoded field of the ZipGAN method, $\mathrm{\phi_{enc}}$, is shown for different downsampling factors, $\mathrm{n_{\Delta}}$.
• Figure 3: 2D slices of instantaneous normalized velocity magnitude and the absolute error of $\phi_{\mathrm{dec,i}}$ fields versus original DNS. The normalized absolute error is given by $\mathrm{\hat{E}=E/\max(E_{\phi_{orig},\phi_{dec,i}},E_{\phi_{orig},\phi_{dec,i}})}$, where $\mathrm{E=\left|\textbf{u}_{\phi_{\mathrm{orig}}} -{\textbf{u}}_{\phi_{\mathrm{dec,i}}}\right|}$ and the subscript $\mathrm{i}$ indicates either the DWT or the ZipGAN data compression methods.
• Figure 4: Comparison of the decoding performance at various compression ratios, $\mathrm{\theta}$, based on the averaged NMSE of the velocity gradient, $\bar{\mathcal{E}} ({|\nabla \mathbf{u}|})$, (bars), and the SSIM between $\mathrm{\phi_{dec}}$ and $\mathrm{\phi_{orig}}$ (solid line). The comparison is made using DWT, ZipCNN, and ZipGAN compression methods on the Re90 dataset (left) and the Re210 dataset (right).
• Figure 5: Number of snapshots ($\mathrm{\# \, snapshots}$) required (bars) and training time ($\mathrm{T_{training}}$) (solid line) needed to achieve the accuracy levels reported in Fig. \ref{['fig:Recap_errors']} (left), with and without the use of transfer learning (TL), on the Re90 dataset.