Machine learning approaches for ray-based ocean acoustic tomography.

Area of Science:

• Oceanography
• Acoustics
• Geophysics

Background:

• Underwater sound propagation relies on nonlinear models linking ocean sound speed profiles (SSPs) to acoustic observations.
• Ocean acoustic tomography (OAT) reconstructs SSP variations using acoustic measurements from multiple source-receiver pairs.

Purpose of the Study:

• To investigate and compare the performance of three OAT methods: model-based, data-driven (deep learning), and a hybrid neural adjoint (NA) method.
• To evaluate these methods using synthetic data for a downward refracting environment with thermocline fluctuations.

Main Methods:

• Model-based OAT: Classical ray-based approach with a linearized forward model.
• Data-driven OAT: Deep learning to directly learn the inverse model.
• Hybrid OAT (NA method): Combines deep learning of the forward model with recursive optimization.

Main Results:

• Hybrid methods can enhance OAT predictions, particularly with favorable sensing configurations and ray coverage.
• The effectiveness of OAT methods is influenced by the dynamics of SSP variations.
• Synthetic SSPs were used to increase training set variability.

Conclusions:

• Merging data-driven and model-based approaches offers potential benefits for OAT.
• The choice of sensing configuration and ray coverage significantly impacts OAT performance.
• The inherent dynamics of sound speed profile variations are critical for robust OAT predictions.