PHYSICS-AIDED DATA-DRIVEN UNDERWATER ACOUSTIC PROPAGATION MODELING

Abstract

The ability to effectively model acoustic propagation is vital for numerous underwater applications. Conventional physics-based propagation models numerically solve the acoustic wave equation. They require full and accurate environmental parameters that may not always be measurable in practice. While classical data-driven machine learning (ML) techniques allow us to predict acoustic fields from data, they are data-hungry and lack extrapolability and interpretability. We design a class of ML algorithms that the physics of acoustic propagation is encoded in the structures of the algorithms. The underlying physical constraint not only enables a data-efficient model, but also offers flexibilities to combine classical ML models and incorporate varying degrees of environmental knowledge, brings interpretability to trained model parameters and generalizes well to permit extrapolation beyond the area where data is collected. We demonstrate the superiority and applicability of proposed hybrid modeling frameworks through simulation studies and a controlled experiment.