Predicting the dynamics of a gas pocket during breaking wave impacts using machine learning

TL;DR

This paper develops a machine learning model to accurately predict the behavior of gas pockets during wave impacts on walls, using experimental data and convolutional LSTM networks.

Contribution

It introduces a novel ML approach trained on experimental wave impact data to predict gas pocket dynamics with high accuracy.

Findings

ML model accurately predicts pressure extremes and oscillation frequency

Experimental data shows gas pocket oscillation depends on initial volume and geometry

Model outperforms traditional models in capturing gas pocket behavior

Abstract

We investigate the feasibility and accuracy of a machine learning model to predict the dynamics of a gas pocket that is formed when a breaking wave impacts on a solid wall. The proposed ML model is based on the convolutional long short-term memory structure and is trained with experimental data. In particular, it takes as input two high-speed camera snapshots before impact and produces as output six scalars that describe the dynamics of the gas pocket. The experiments are performed in a wave flume, where we use solitons -- in combination with a bathymetry profile -- to generate wave breaking close to a solid wall which is instrumented with dynamic pressure sensors. By varying the water depth $h_\ell$ and the parameter $\alpha = A/h_\ell$, where $A$ is the soliton wave amplitude, we are able to generate a family of unique breaking waves with different gas pocket sizes and wave kinematics. In this so-called phase space of wave generation ($h_\ell$, $\alpha$), we perform experiments on 67 different wave states that form our dataset. Experimentally, we find that the frequency of oscillation of the gas pocket can be attributed to the initial volume of gas plus a geometric correction and that the maximum and minimum pressures are qualitatively well captured by the one-dimensional Bagnold model. In terms of the ML model, we compare its performance to the experimental data and find that the model quantitatively reproduces the trends found in the experiments -- in particular for the maximum and minimum pressure in the gas pocket and the frequency of oscillation.