Investigating Machine Learning Super-Resolution for Spatiotemporal Experimental Data

With the advent of high-performance computing capabilities, high fidelity turbulent flow data has become increasingly accessible, allowing for advancements in our understanding of turbulent flow physics. However, simulations often require expensive computations. Experimental methods such as PIV come with certainty of physical accuracy and inexpensive data acquisition given a proper setup, but the acquisition of spatiotemporally resolved data remains a challenge due to physical constraints, especially with increasing Reynolds number. To balance the physical constraints of PIV and the computational expense of simulations, we explore the use of machine learning super-resolution to augment under-resolved experimental spatiotemporal flow data. We consider convolutional neural networks and transformer-based model architectures, which have both shown success in improving spatiotemporal resolution of turbulent flow data. We explore various pooling methods and discuss their suitability for the generation of training datasets. We validate our methods by comparison with PIV data of varying resolution in both space and time, analyzing the effects of model parameters and pooling methods for their ability to reconstruct multiscale turbulent flow physics.