High-frequency underwater acoustic field modeling is challenging for physics-informed neural networks (PINNs) due to the strongly oscillatory nature of acoustic pressure fields. Building upon the OceanPINN framework, this paper proposes a separable-variable OceanPINN with learnable spectral expansions for two-dimensional acoustic field prediction in high-frequency environments. The acoustic envelope is represented as a separable expansion of range-dependent and depth-dependent basis functions learned by one-dimensional spectral neural networks, with trainable expansion coefficients. Analytical derivatives of the trigonometric bases enable efficient and stable enforcement of the Helmholtz equation. Numerical results demonstrate improved accuracy and generalization over conventional OceanPINN under limited data conditions.
Ding et al. (Mon,) studied this question.