Fluid Dynamics Measurements using Deep Learning Based Optical Flow Models

Particle image velocimetry (PIV) is a gold-standard optical technique for quantifying velocity fields from tracer particle images; however, its post-processing remains highly sensitive to correlation parameters and typically requires large numbers of image pairs to achieve converged results, limiting its use in time-resolved and unsteady flow measurements. This limits PIV application to rapidly fluctuating, unsteady flows especially when the image quality is also challenged due to image artifacts. This study addresses two key challenges in PIV post-processing: reducing parameter sensitivity and identifying the minimum number of frames required for reliable velocity estimation across different flow regimes. To this end, three optical flow–based approaches—Dense Optical Flow (DOF), Spatial Pyramid Network (SPyNet), and Recurrent All-Pairs Field Transforms (RAFT)—are systematically evaluated as alternatives or complements to classical PIV. The methods are tested on synthetically generated PIV datasets with known ground truth and on experimental jet flow measurements spanning laminar and turbulent regimes at Reynolds numbers between 1100 and 2200. For synthetic translational particle motion, DOF achieved a best-case MAPE of 5.9%, while SPyNet yielded dataset-dependent errors ranging from 2.6% to 5.5%, both outperforming standard PIV. In rotational synthetic flows, DOF provided the lowest errors (1.5–3% for angular velocities of 2°–5°), comparable to classical PIV, whereas SPyNet exhibited higher errors (4.4–5.9%). RAFT showed limited performance on synthetic translational motion (34–37% MAPE), while domain-specific fine-tuning improved its accuracy mainly for rotational cases, reducing the error to approximately 30%. In turbulent flow conditions, DOF consistently captured temporal velocity fluctuations with a percentage standard deviation of approximately 10% using as few as four frames, whereas SPyNet required larger frame counts (up to 15 frames) to achieve comparable fluctuation statistics at higher Reynolds numbers. These results demonstrate that optical flow–based methods can substantially reduce frame requirements for time-resolved PIV analysis, but their effectiveness is strongly dependent on flow regime and data characteristics. Rather than a universal replacement for PIV, this study provides a regime-aware framework for selecting suitable optical flow models to enable efficient and time-resolved fluid dynamics measurements.