Abstract
Abstract
Current optical flow-based neural networks for particle image velocimetry (PIV) are largely trained on synthetic datasets emulating real-world scenarios. While synthetic datasets provide greater control and variation than what can be achieved using experimental datasets for supervised learning, it requires a deeper understanding of how or what factors dictate the learning behaviors of deep neural networks for PIV. In this study, we investigate the performance of the recurrent all-pairs field transforms-PIV (RAFTs-PIV) network, the current state-of-the-art deep learning architecture for PIV, by testing it on unseen experimentally generated datasets. The results from RAFT-PIV are compared with a conventional cross-correlation-based method, Adaptive PIV. The experimental PIV datasets were generated for a typical scenario of flow past a circular cylinder in a rectangular channel. These test datasets encompassed variations in particle diameters, particle seeding densities, and flow speeds, all falling within the parameter range used for training RAFT-PIV. We also explore how different image pre-processing techniques can impact and potentially enhance the performance of RAFT-PIV on real-world datasets. Thorough testing with real-world experimental PIV datasets reveals the resilience of the optical flow-based method’s variations to PIV hyperparameters, in contrast to the conventional PIV technique. The ensemble-averaged root mean squared errors between the RAFT-PIV and Adaptive PIV estimations generally range between 0.5–2 (px) and show a slight reduction as particle densities increase or Reynolds numbers decrease. Furthermore, findings indicate that employing image pre-processing techniques to enhance input particle image quality does not improve RAFT-PIV predictions; instead, it incurs higher computational costs and impacts estimations of small-scale structures.