Full Program »
SUPER-RESOLUTION ANALYSIS WITH MACHINE LEARNING FOR LOW-RESOLUTION TURBULENT FLOW DATA
We present machine learning-based super-resolution techniques to reconstruct subgrid-scale turbulent flow fields from low-resolution (LR) data. Two-dimensional turbulent flow data obtained from direct numerical simulation is con- sidered, with which LR data sets are generated by max- pooling operation. We construct machine learning (ML) models based on the convolutional neural networks (CNN) to predict high-resolution (HR) flow fields from LR data. Resolved flow field by our machine learned method called the Downsampled Skip-Connection Multi-Scale (DSC-MS) model shows agreement between the reconstructed flow field and the reference solution in terms of its flow field and kinetic energy cascade. Towards the end of the paper, we offer discussions on the applicability of the current ap- proach for general flow field data obtained from numerical simulations and experimental measurements.