Robust Flow Field Reconstruction Using PINN for 3D Lagrangian Particle Tracking

Abstract

Physics-Informed Neural Networks (PINNs) effectively reconstruct fluid flow fields from sparse Lagrangian Particle Tracking (LPT) data by embedding physical laws directly into neural network training. This study investigates the influence of experimental parameters such as particle density on PINN reconstruction performance using synthetic (DNS-based HIT and turbulent channel flows) and experimental turbulent boundary layer (TBL) datasets. Results demonstrate PINNs’ robustness across varying seeding densities, with notably superior performance in TBL cases compared to HIT cases. Optimization techniques, particularly the SOAP optimizer, significantly enhance convergence speed and accuracy, highlighting PINNs' potential for reliable fluid flow reconstruction from limited experimental data.