Physics-informed neural networks for gravity currents reconstruction from limited data

TL;DR

This paper explores the use of physics-informed neural networks (PINNs) to reconstruct 3D unsteady gravity currents from limited data, demonstrating their effectiveness through numerical experiments and benchmarking different measurement techniques.

Contribution

The study introduces a PINN-based framework for gravity current reconstruction and evaluates its performance with various limited data scenarios, including experimental measurement mimicking techniques.

Findings

PINNs successfully reconstruct gravity currents from limited data.

Spatially averaged density measurements via LAT are effective for training.

An optimal setup balances implementation complexity and reconstruction accuracy.

Abstract

The present work investigates the use of physics-informed neural networks (PINNs) for the 3D reconstruction of unsteady gravity currents from limited data. In the PINN context, the flow fields are reconstructed by training a neural network whose objective function penalizes the mismatch between the network predictions and the observed data and embeds the underlying equations using automatic differentiation. This study relies on a high-fidelity numerical experiment of the canonical lock-exchange configuration. This allows us to benchmark quantitatively the PINNs reconstruction capabilities on several training databases that mimic state-of-the-art experimental measurement techniques for density and velocity. Notably, spatially averaged density measurements by light attenuation technique (LAT) are employed for the training procedure. An optimal experimental setup for flow reconstruction by PINNs is proposed according to two criteria : the implementation complexity and the accuracy of the inferred fields.