A robust super-resolution reconstruction model of turbulent flow data based on deep learning

Abstract

A new super-resolution model, namely the turbulence volumetric super-resolution (TVSR) model, is developed based on convolutional neural network (CNN) to reconstruct three-dimensional high-resolution turbulent flow field data from low-resolution data. Direct numerical simulation (DNS) and corresponding filtered DNS (FDNS) data of homogeneous isotropic turbulence at various Reynolds numbers are used to train the TVSR model. The proposed model is a modification of Liu et al. (2020), aiming to provide an improved generalization capability of the super-resolution model. For this purpose, we propose a patchwise training strategy in consideration of the property of turbulence that the velocity correlation between two points diminishes as the separation becomes sufficiently large. Furthermore, data at various Reynolds numbers are combined together to train the model. In comparison with existing models, the present TVSR model shows a better generalization capability in two aspects. First, the TVSR model trained using data at low Reynolds numbers is found robust and accurate in the super-resolution reconstructions of flow fields at higher Reynolds numbers. Second, although only DNS data are used for training, the TVSR model is also robust in reconstructing high-resolution flow fields from low-resolution data obtained from large-eddy simulation (LES). This feature of the TVSR model provides a new access to obtain turbulent motions at unresolved scales in LES studies of turbulent flows.