PINN-enhanced particle tracking velocimetry

Particle tracking velocimetry (PTV), capable of providing a non-intrusive way for global velocity measurement, has been playing an important role in experimental fluid mechanics. However, due to the limitation of the particle matching algorithms, the resulting velocity field by PTV is generally sparse and noisy. In this work, we propose to use physics-informed neural networks (PINNs) to deal with this problem. The PINN-enhanced PTV approach assimilates the particle tracking vectors and the governing equations of the investigated flow, and then returns a temporally- and spatially-continuous flow field. In this context, the resulting velocity field is physically-interpretable and other physical quantities, including pressure and vorticity, can be inferred simultaneously. Moreover, we demonstrate that by using the PINN-enhanced PTV, we can downsample the velocity measurements in space and time to an extremely sparse level. We also show that with the boundary conditions imposed, the PINN-enhanced PTV is able to extend the flow field well beyond the observation domain. The proposed method is evaluated on various flow topologies (horseshoe vortex flow, wake flow and jet flow) and different experimental setups (2D PTV and Tomo-PTV).  We demonstrate that the PINN-enhanced method is promising to become a standard way for processing the PTV data.