Stable adaptive training for physics-informed neural networks in acoustic wave propagation

Area of Science:

• Computational physics
• Acoustics
• Machine learning

Background:

• Physics-informed neural networks (PINNs) show promise for scientific simulations.
• Underwater acoustics simulations face challenges with large domains and convergence issues.
• Existing PINN methods struggle with computational domain size and trivial solution convergence.

Purpose of the Study:

• To enhance the application of PINNs in underwater acoustics.
• To address challenges of large computational domains and convergence to trivial solutions in PINN training.
• To develop a strategy for more stable and accurate acoustic wave propagation simulations.

Main Methods:

• Combining adaptive domain sampling with absorbing boundary conditions for PINN training.
• Dynamically focusing computational effort on regions of localized acoustic energy.
• Utilizing absorbing boundaries to stabilize the training process.

Main Results:

• The proposed method significantly improves the stability of PINN training.
• Enhanced convergence rates were observed in numerical experiments.
• The strategy leads to more accurate and reliable wave propagation simulations.

Conclusions:

• Adaptive domain sampling and absorbing boundary conditions effectively address PINN limitations in underwater acoustics.
• The developed strategy offers a robust approach for complex acoustic simulations.
• This work paves the way for more efficient and precise underwater acoustic modeling using PINNs.