Physics-guided machine learning reconstruction of modal structure in the seabed characterization experiments

Abstract

Accurate extraction of normal modes from underwater acoustic data is relevant for seabed characterization and geoacoustic analysis. In this context, efficiently predicting these depth-dependent mode structures is critical for understanding the acoustic field. Standard signal processing techniques for extracting these features can be computationally intensive or sensitive to noise, limiting their robustness in practical applications. This work introduces a physics-guided machine learning framework designed to predict depth-dependent mode shapes in a range-independent waveguide. The model is developed using a synthetic dataset generated with a normal mode model, parameterized with experimentally measured sound speed profiles from the Seabed Characterization Experiments (SBCEX) in 2017 and 2022. By learning the physical relationships defined by the waveguide, the framework provides a data-driven approach to characterize modal structures. The motivation is to enable modal structure estimation from sparse vertical sampling typical of SBCEX (8 hydrophones), where conventional extraction is unstable. We evaluate the model's performance in the low-frequency regime, demonstrating its potential as a robust, automated methodology for modal analysis in realistic ocean environments, where the model achieved correlation above 0.9 (averaged across modes 1–5) for frequencies below 200 Hz. [Work supported by ONR code 322.]