A neural network-based model for magnetic particle tracking and its application in the study of fluidization

The recently developed magnetic particle tracking (MPT) technology provides a new tool for flow diagnosis in fully opaque environments like those in fluidized bed and sediment transport. The MPT can measure both translation and rotation of a particle, which is crucial to study the particle dynamics. Current reconstruction algorithms in MPT include semi-analytical solver, optimization, and Kalman filter, among others. These methods have certain limitations. For instance, the analytical solver can hardly be extended to multi-particle systems, and the optimization is time-consuming. In this study, a neural network (NN) based model is developed to reconstruct the particle trajectory and orientation. The network architecture uses the multi-layer perceptron for reconstruction and gated recurrent unit for denoising. After initial training, the network model can provide particle trajectories in real time. More importantly, the NN-based model can simultaneously reconstruct trajectories of several particles, which is a big advantage compared to the existing methods. This progress enables us to study particle collisions, two-point correlations, and other statistical properties. This system has been used to study the dynamics of particles in a fluidized bed.