Quantitative Assessment of PINN Inference on Experimental Data for Gravity Currents Flows

TL;DR

This paper demonstrates that Physics Informed Neural Networks (PINNs) can accurately infer velocity and pressure fields from experimental data on gravity currents, showing robustness to noise and potential for real-world engineering applications.

Contribution

The paper applies PINNs to real experimental gravity current data, validating their effectiveness in inferring physical fields from sparse, noisy measurements.

Findings

PINNs accurately infer velocity fields from LAT data.

PINNs show robustness to noise in experimental data.

Quantitative comparison with PIV confirms model accuracy.

Abstract

In this paper, we apply Physics Informed Neural Networks (PINNs) to infer velocity and pressure field from Light Attenuation Technique (LAT) measurements for gravity current induced by lock-exchange. In a PINN model, physical laws are embedded in the loss function of a neural network, such that the model fits the training data but is also constrained to reduce the residuals of the governing equations. PINNs are able to solve ill-posed inverse problems training on sparse and noisy data, and therefore can be applied to real engineering applications. The noise robustness of PINNs and the model parameters are investigated in a 2 dimensions toy case on a lock-exchange configuration, employing synthetic data. Then we train a PINN with experimental LAT measurements and quantitatively compare the velocity fields inferred to Particle Image Velocimetry (PIV) measurements performed simultaneously on the same experiment.The results state that accurate and useful quantities can be derived from a PINN model trained on real experimental data which is encouraging for a better description of gravity currents.