Neural Network Approach to Traditional Particle Image Velocimetry

Particle Image Velocimetry (PIV) has become a standard tool in fluid dynamics to achieve planar and volumetric velocity measurements. Since its inception, calculation of velocity vectors have most commonly been achieved by tracking particles seeding a particular flow with spatial cross-correlation methods to calculate particle shifts in a certain time window. While this method of velocity vector calculation can be reliable in planar applications, measurement uncertainty, algorithm robustness, and scientist in the loop decision making have remained areas of concern. This work presents an outline for a PIV neural network architecture. Rather than training the neural network to identify and exploit physics, this work seeks to substitute a neural network into the contemporary PIV pipeline as a robust alternative to spatial cross correlations. The neural network is trained to track particle shifts over sub-regions of a particle image field with the added benefit of a more holistic treatment of uncertainty via probability.