A parabolic equation-based physics-informed machine learning method for underwater sound propagation modelinga)

Area of Science:

• Ocean acoustics
• Underwater physics
• Machine learning applications

Background:

• Traditional underwater sound propagation models face computational and adaptability limits.
• Data-driven methods lack physical constraints and require large datasets.
• Physics-Informed Neural Networks struggle with accurate long-range extrapolation.

Purpose of the Study:

• To develop a physics-informed machine learning method for accurate long-range underwater sound propagation modeling.
• To enhance the stability and reduce error accumulation in acoustic propagation predictions.
• To improve underwater environmental monitoring, communication, and localization.

Main Methods:

• Proposed U-PARANET, a physics-informed machine learning model.
• Incorporated the parabolic equation as a hard constraint within the neural network architecture.
• Leveraged the parabolic equation's recursive structure for enhanced stability.

Main Results:

• U-PARANET accurately predicted transmission loss and phase structures in simulated and experimental data.
• Achieved low mean absolute errors for transmission loss: 1.40 dB (ideal), 1.06 dB (simulated SWellEx-96), and 2.87 dB (experimental SWellEx-96).
• Demonstrated good agreement in spatial field patterns and robust long-range extrapolation.

Conclusions:

• U-PARANET exhibits excellent long-range underwater sound propagation modeling capabilities.
• The method shows robust extrapolation in challenging, realistic ocean environments.
• U-PARANET offers a promising advancement for oceanographic applications.