Research On Turbulent Super-resolution Algorithm Based On Deep Learning

Turbulence is a widely existing flow form in nature and has extensive applications in various fields.However,the nonlinearity,high dimensionality,multiscale nature,randomness,and uncertainty of turbulence pose high requirements for the simulation,acquisition,processing,and analysis of turbulence data.The resolution of flow field data is an important factor affecting the accuracy of turbulence field analysis.Obtaining high-resolution turbulence data has become a hot issue in the field of turbulence analysis and processing.However,acquiring high-resolution turbulence often requires a large amount of resources.Experimentally,expensive and precise laboratory equipment is needed,and computationally,complex partial differential equations need to be solved repeatedly,requiring a large amount of computational resources.Therefore,how to reconstruct high-resolution details from low-resolution simulation results is a challenging and practical problem.Although super-resolution reconstruction for turbulence flow field is an effective way to obtain its details,traditional interpolation methods are difficult to reconstruct small-scale structures in turbulence,and their results are too smooth.Some CNN-based methods have insufficient accuracy or cannot restore the physical properties of flow fields.This paper focuses on the deep learning method of turbulence super-resolution,and the main research contents are as follows:1.A supervised Super-Resolution Transformer for Turbulence(SRTT)model is proposed for acquiring high-quality high-resolution turbulent flow fields.It is capable of restoring high-resolution images from single-frame low-resolution images and effectively eliminating problems such as blurring in low-resolution turbulence data.With self-attention mechanism,the model can effectively capture long-range dependency between each pixel,thus better reconstructing the global structure of turbulence.This enables the algorithm to maintain the quality of high-resolution results at different resolutions.2.An improved SRUTT model is proposed for the super-resolution turbulence model,which used U-Net structure for skip connections to preserve more spatial information by connecting different feature maps of the flow field to the output of the Transformer structure.The model is validated on turbulent channel flow and comprehensive comparative analysis is performed on the reconstructed high-resolution turbulent flow fields.Experimental results showed that even for anisotropic flows with large differences in velocity in the three directions and with different physical properties in each region of the boundary layer,the model still reconstruct a good high-resolution turbulent channel flow field while preserving good physical properties such as Reynolds stresses,kinetic energy spectra,etc.In summary,the deep learning-based turbulence super-resolution algorithm proposed in this paper can effectively use low-resolution turbulence flow field data to obtain corresponding high-resolution flow fields.It can effectively reuse some existing low-resolution data and avoid the high complexity and resource consumption of obtaining high-resolution turbulent flow fields in traditional fluid mechanics methods.Through in-depth exploration of the principles and applications of turbulence super-resolution algorithms,it is expected to provide more accurate tools and methods for turbulence research and promote the continuous development and progress of the turbulence research field.