Three-dimensional super-resolution reconstruction of turbulent flow using 3D-ESRGAN with random sampling strategy

This study introduces a deep learning framework that uses an enhanced three-dimensional super-resolution generative adversarial network (3D-ESRGAN) to reconstruct high-resolution turbulent flow fields from low-resolution data. To minimize the reliance on complete datasets during training, a random s...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Computers & fluids Ročník 305; s. 106890
Hlavní autoři: Yu, Linqi, Chen, Yanyun, Yousif, Mustafa Z., Lim, Hee-Chang
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 30.01.2026
Témata:
ISSN:0045-7930
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:This study introduces a deep learning framework that uses an enhanced three-dimensional super-resolution generative adversarial network (3D-ESRGAN) to reconstruct high-resolution turbulent flow fields from low-resolution data. To minimize the reliance on complete datasets during training, a random sampling strategy is used. In each training epoch, approximately 1 % of spatial points are randomly chosen from both the predicted and ground-truth fields for loss computation. This method mimics sparse sensor measurements in real-world experiments, allowing the model to learn accurate mappings based on limited spatial observations. Furthermore, physics-guided loss functions are incorporated to ensure consistency with the underlying physical laws, thereby improving the reliability of the reconstructed flow fields. Subsequently, the framework is tested on two typical flow scenarios: (1) flow over a finite wall-mounted square cylinder at Red=500, and (2) fully developed turbulent channel flow at Reτ=180 and 500. Both scenarios are generated using direct numerical simulation (DNS). Finally, the results are presented and analyzed both qualitatively and quantitatively. The qualitative findings show that the reconstructed fields effectively restore the key vortex structures and turbulence features absent in the coarse data. Quantitative comparisons with the ground-truth data confirm high accuracy in terms of velocity profiles, Reynolds stresses, probability density functions (PDFs), and energy spectra. Additionally, the relative error regarding the streamwise velocity magnitude is calculated, and all cases show a low error rate of approximately 5 %. In summary, the findings confirm the efficacy of combining GAN-based super-resolution with a random sampling strategy for accurate and data-efficient 3D turbulence reconstruction. The results suggest that the proposed framework can be successfully applied in real-world scenarios where only sparse measurements are accessible.
ISSN:0045-7930
DOI:10.1016/j.compfluid.2025.106890