Machine Learning Aided Flow Field Reconstruction from Sparse and Noisy Particle Measurements

Cavitation can cause significant erosion to the blades of marine propellers and increase the vibration and noise of ships. Therefore, it is important to better understand cavitation inception corresponding to the first appearance of rapid bubble growth. Cavitation inception is driven by the minimum pressure in the flow, which can be difficult to accurately measure in experiments due to the sparse distribution of particles around the vortex cores. In this work, we apply a physics-informed deep learning (PIDL) method to reconstruct the complete flow field from sparse particle tracks with a focus on accurately assimilating the minimum pressure on a fine grid domain from sparse particle trajectories with velocity information. We tested the performance of the deep learning model using a series of classical vortex models in 2D and 3D. Synthetic PTV data are generated using the analytical solutions of the velocities, those sparse PTV data with velocity information are trained by the DL model, and then the model reconstructs the high-resolution flow field within a fixed domain. Preliminary results show that the PIDL model can successfully reconstruct the fine flow field of both velocity and pressure with an observation rate less than 1%. This work also investigates the influence of particle density and distributions to the performance of the model. The final results will include the performance of the proposed method to noisy data and application to data from high-fidelity numerical simulations of vortex flows.