A deep-learning super-resolution reconstruction model of turbulent reacting flow

Abstract

Super-resolution reconstruction using deep-learning techniques is a valuable tool for enhancing computational and experimental visualization of flow fields to recover detailed features from coarse results. In this study, we propose the Pyramid Res U Super Resolution (PRUSR) network as a flow reconstruction model that upgrades the basic architecture of the Downsampled Skip-Connection Multi-Scale (DSC/MS) model with the pre-activated residual block (PRB). To improve the model accuracy, we also implement a new loss function called Mean Square Structure Similarity (MSSSIM), which combines the Mean Square Error (MSE) and Structure Similarity (SSIM). We test the model on a practical problem of visualizing the three-dimensional turbulent jet flame and find that the PRUSR model accurately reconstructs the super-resolution flow field, particularly in large downsampling ratio conditions, outperforming the original DSC/MS model and traditional interpolation algorithms such as bilinear and bicubic even with fewer input data. Additionally, the PRUSR model is capable of reconstructing sub-grid turbulent flow and flame thickness by inferring high-frequency information in the inertial subrange, making it a powerful post-processing tool for turbulent flow visualization and analysis.