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

Physics-informed neural networks (PINNs) have emerged as a promising tool for simulating various phenomena. However, their application in underwater acoustics remains challenging, primarily due to the need to sample large computational domains and to convergence to trivial solutions. This study presents a strategy to address these issues by combining adaptive domain sampling with absorbing boundary conditions. The adaptive sampler dynamically focuses computational effort on regions where the acoustic energy is localized, while the absorbing boundaries perform training stabilization. Numerical experiments show that our method improves the stability and convergence of PINN training, leading to more accurate and reliable wave propagation simulations.