Leveraging sound speed dynamics and generative deep learning for ray-based ocean acoustic tomography.

Area of Science:

• Oceanography
• Acoustics
• Machine Learning
• Geophysics

Background:

• Ocean acoustic tomography (OAT) is an inverse problem to estimate sound speed profiles (SSP).
• Traditional OAT methods often face ill-posed challenges, limiting accuracy in SSP estimation.
• Accurate SSPs are crucial for underwater navigation, communication, and marine research.

Purpose of the Study:

• To introduce a generative deep learning framework for ray-based OAT.
• To enhance the accuracy and robustness of SSP estimation in ocean environments.
• To overcome the ill-posed nature of conventional OAT formulations.

Main Methods:

• Utilized a generative deep learning framework incorporating a variational autoencoder.
• Employed a linear dynamical model for robust low-dimensional parametrization of SSP variations.
• Tested the framework with simulated SSP variations from a regional ocean model and various transducer setups.

Main Results:

• The deep learning framework significantly outperformed conventional linear least squares OAT formulations.
• Demonstrated robust performance in estimating SSPs under varying ocean conditions and transducer configurations.
• Successfully addressed the ill-posed nature of OAT using advanced machine learning regularization.

Conclusions:

• The proposed generative deep learning framework offers a powerful advancement for ocean acoustic tomography.
• This approach provides more accurate and reliable sound speed profile estimations.
• The findings have significant implications for improving underwater acoustic sensing and oceanographic studies.