On the accuracy of data assimilation algorithms for dense flow fields reconstructions

Kurzfassung

In this study, we explore the application of Physics Informed Neural Networks (PINNs) in approximating 3D flow fields. Synthetic datasets with varying particle concentrations from a subsampled DNS domain of forced isotropic turbulence are provided to PINNs. Through training PINNs, we derive fully reconstructed velocity and pressure fields for all cases and compare them to the ground truth. It is found that PINNs generally well approximate velocity and pressure, but there are still limitations in correctly predicting high-frequency regions.