Deep learning-based super-resolution reconstruction model for cavitation images in cryogenic flow fields

Abstract

Obtaining cavitation images of cryogenic fluid flow fields with high temporal and spatial resolution has important scientific value for revealing the mechanism of cryogenic cavitation. This study develops a super-resolution (SR) reconstruction model for cryogenic cavitation flow fields based on deep learning architectures (SRCNN, SRResNet, SRGAN, SwinIR) trained on the datasets obtained from Venturi tube liquid nitrogen (LN₂) cavitation flows. Results show that the proposed model achieves ×4 and ×8 spatial resolution enhancement of the cavitation images, thus restoring flow field details well. By analyzing the performance of the model under different pressure ratios P_r inside and outside the dataset, and combining it with the convergent-divergent nozzle LN₂ flow field data for verification, it is found that the proposed model can accurately recover fine cavitation flow features from low-resolution inputs while maintaining the temporal coherence and exhibiting good generalization across operational conditions. Quantitative analysis shows that the model based on SwinIR performs better in peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and image standard deviation metrics, maintaining robustness even in severe cavitation information loss. The proposed method provides an effective SR analysis tool for cavitation experiments and has important application value for exploring the cryogenic cavitation mechanism.