A broadband modeling method for range-independent underwater acoustic channels using physics-informed neural networks

Area of Science:

• Underwater acoustics
• Computational physics
• Machine learning

Background:

• Accurate broadband modeling of underwater acoustic channels is crucial for applications like detection, localization, and communication.
• Conventional methods (FEM, FDM, BEM) are computationally intensive, requiring single-frequency calculations and limiting broadband analysis.
• Existing methods face significant time challenges for broadband modeling due to repeated frequency computations.

Purpose of the Study:

• To develop a rapid broadband modeling approach for underwater acoustic channels.
• To overcome the computational limitations of conventional single-frequency methods.
• To enhance the efficiency of underwater acoustic channel response prediction.

Main Methods:

• Utilizing physics-informed neural networks (PINNs) for rapid broadband modeling.
• Integrating the modal equation of normal modes as a regularization term in the neural network's loss function.
• Applying the method in range-independent underwater environments with a liquid semi-infinite seabed.

Main Results:

• The PINN method achieves rapid broadband modeling using sparse frequency sampling points.
• Accurate prediction of the underwater acoustic channel response from 100 to 300 Hz.
• A 25-fold improvement in computational speed compared to KRAKEN at 20 km propagation distance.
• Maintained a mean absolute error of 0.15 dB for the acoustic channel response.

Conclusions:

• Physics-informed neural networks offer a computationally efficient solution for broadband underwater acoustic channel modeling.
• The proposed method significantly reduces computation time while preserving accuracy.
• This approach has strong potential for advancing underwater acoustic detection, localization, and communication systems.