An accelerated coupled normal-mode model based on physics-informed neural networksa).

Area of Science:

• Ocean acoustics
• Computational physics
• Machine learning applications

Background:

• Traditional coupled normal-mode models for underwater sound propagation are computationally expensive due to repeated modal Sturm-Liouville problem solutions.
• Range-dependent environments in underwater acoustics present significant simulation challenges.

Purpose of the Study:

• To develop an accelerated coupled normal-mode model using physics-informed neural networks (PINNs) for efficient underwater sound propagation simulation.
• To improve computational efficiency and maintain accuracy in simulating sound propagation in complex, range-dependent shallow water environments.

Main Methods:

• Utilized Empirical Orthogonal Function (EOF) analysis to compress sound-speed profiles.
• Developed a dual-branch PINN to predict horizontal wavenumbers and modal depth functions.
• Integrated data-driven and physics-based constraints (modal equation, boundary conditions) into the PINN training loss.

Main Results:

• Achieved a mode-averaged relative error of 1.52 × 10-3% for wavenumbers and 6.7% for depth functions on the 2015 Shallow Water Sound Fluctuation dataset.
• Reduced online modal-solving time by 85% compared to finite difference solvers, outperforming first-order modal perturbation theory.
• Demonstrated transmission loss errors within 3 dB and complex pressure normalized mean squared errors on the order of 10-5 against KRAKENC simulations.

Conclusions:

• The proposed physics-informed neural network framework offers a practical and computationally efficient solution for underwater sound propagation in range-dependent shallow water environments.
• The model accurately simulates sound propagation considering combined sound-speed and bathymetric variations.
• This approach significantly enhances the speed of acoustic simulations without compromising accuracy.