The process of flows carrying particles is highly complex, traditionally tackled by solving the Navier-Stokes equations. Although different numerical and experimental techniques have been developed, these approaches demand a deep understanding of the underlying physics and blackare frequently associated with high computational costs. Machine learning offers a novel alternative, learning predictive patterns directly from data, thus bypassing the need for explicit physical modeling. Nonetheless, pure data-driven methods can sometimes lack interpretability and physical consistency. By integrating physics principles into machine learning, this gap can be bridged and the above problems can be solved. In this context, we have proposed TrajectoryFlowNet for flow and particle tracking. Our approach combines the flexibility of data-driven learning with the rigorousness of physics-based constraints, aiming to achieve both accuracy and efficiency. The salient features of our model include its ability to handle complex flow patterns with moving boundaries, predict the trajectories of all particles in the domain, and ensure physical consistency throughout the predictions based only on sparse trajectories. To validate our method, we have conducted several numerical and experimental cases across a range of flow scenarios. These experiments demonstrate the model's effectiveness in capturing the intricate dynamics of particle-laden flows, advancing precise particle tracking and flow field inversion in various real-world problems.
Wan et al. (Sun,) studied this question.