Identification of Settling Velocity with Physics Informed Neural Networks For Sediment Laden Flows

TL;DR

This paper demonstrates how Physics-Informed Neural Networks can accurately recover the settling velocity and reconstruct flow fields in sediment-laden flows, even with noisy data, advancing fluid dynamics modeling.

Contribution

It introduces a novel application of PINNs to sedimentation flows, effectively inferring unknown parameters and flow fields across different regimes with noisy data.

Findings

PINNs successfully recover settling velocity in sediment-laden flows.

The model accurately reconstructs flow fields compared to high-fidelity simulations.

PINNs show robustness to noisy data, maintaining inference quality.

Abstract

Physics-Informed Neural Networks (PINNs) have shown great potential in the context of fluid dynamics simulations, particularly in reconstructing flow fields and identifying key parameters. In this study, we explore the application of PINNs to recover the dimensionless settling velocity for sedimentation flow. The flow involves sediment-laden fresh water overlying salt water, which is described by Navier-Stokes equations coupled with sediment concentration and salinity transport equations. Two cases are investigated: one where the training data contains the salinity and sediment concentration fields, and another where it contains the velocity field. For both cases, we investigate several flow regimes and show that the model is capable of inferring the unknown parameter and reconstructing the hydrodynamic field of the flow. The quality of the model inference is assessed by comparing it with numerical simulations from a high-fidelity semi-Lagrangian solver. We demonstrate the model's robustness to noise by training it with data corrupted by noise of varying magnitudes, highlighting the potential of PINNs for real-world applications.