Spatial domain decomposition-based physics-informed neural networks for practical acoustic propagation estimation...

Area of Science:

• Ocean acoustics
• Computational physics
• Machine learning

Background:

• Acoustic propagation modeling is crucial for oceanic applications but challenged by unknown ocean dynamics.
• Traditional physics-based models require extensive environmental data, while data-driven methods face data collection hurdles.
• Existing physics-informed neural networks (PINNs) offer integration but have limited representational power.

Purpose of the Study:

• To introduce a novel framework, spatial domain decomposition-based physics-informed neural networks (SPINNs), to enhance acoustic propagation estimation.
• To improve the learning capacity of neural networks in spatially dependent oceanic scenarios using incomplete physics and noisy datasets.
• To address the limitations of current methods in practical acoustic propagation modeling.

Main Methods:

• Development of the SPINNs framework, integrating spatial domain decomposition with PINNs.
• Training SPINNs using incomplete prior physics knowledge and noisy datasets.
• Experimental validation of SPINNs against traditional PINNs for acoustic propagation estimation.

Main Results:

• SPINNs demonstrate superior performance over standard PINNs in practical acoustic propagation estimation.
• The framework effectively learns from spatially dependent oceanic scenarios with limited and biased physics information.
• SPINNs show enhanced capacity to learn from noisy datasets compared to existing approaches.

Conclusions:

• SPINNs offer a promising advancement for accurate acoustic propagation modeling in complex, real-world ocean environments.
• The proposed method effectively handles data scarcity and uncertainty in oceanic scenarios.
• This work paves the way for improved practical applications in underwater acoustics and further research expansion.