Spectrum-adaptive physics-informed neural network for rapid ocean acoustic field prediction.

Area of Science:

• Ocean acoustics
• Computational modeling
• Machine learning

Background:

• Physics-informed neural networks (PINNs) face slow convergence challenges in ocean acoustics.
• Efficient modeling of underwater acoustic wave propagation is critical.

Purpose of the Study:

• To develop an efficient modeling approach for ocean acoustics.
• To improve the convergence speed and accuracy of physics-informed neural networks.

Main Methods:

• Proposed a spectrum-adaptive physics-informed neural network (SA-PINN) based on OceanPINN.
• Estimated effective physical bandwidths to determine direction-dependent cutoff frequencies.
• Calibrated sinusoidal representation network hyperparameters for optimal spectral capacity.

Main Results:

• SA-PINN demonstrated state-of-the-art convergence speed compared to standard OceanPINN.
• Achieved significant reductions in computational costs.
• Enhanced modeling accuracy validated by simulations and experimental data.

Conclusions:

• SA-PINN offers a superior initialization strategy for ocean acoustics modeling.
• The proposed method significantly improves efficiency and accuracy in wavefield simulations.
• SA-PINN effectively addresses the convergence limitations of traditional PINNs in this domain.