Ocean acoustic field prediction method based on dual-branch physics-informed neural network under insufficient environmental parameters

Accurate computation of the ocean acoustic pressure field is crucial for underwater acoustic applications. Traditional numerical models depend on comprehensive environmental priors, such as full-depth sound speed profiles (SSPs) and bottom geoacoustic parameters. To address challenges posed by uncertain bottom geoacoustic parameters and depth-truncated SSPs, this study presents a Dual-Branch Physics-Informed Neural Network (DB-PINN) to predict the acoustic pressure field while simultaneously reconstructing the full-depth SSP. The model comprises two parallel branches, predicting the acoustic pressure envelope and the SSP, respectively, utilizing only sparse acoustic pressure observations and shallow-water SSP data. Joint optimization with the wave-equation residual as a physical constraint enables accurate full-depth acoustic-field prediction. Incorporating an envelope representation of the complex acoustic pressure mitigates spectral bias, addressing neural networks’ difficulty in learning high-frequency spatial structures. Validated using measured data from the SWellEx-96 field experiment, DB-PINN extrapolates deep-water SSP behavior and predicts the full-depth complex acoustic field structure, even without explicit prior knowledge of bottom geoacoustic parameters, while using SSP observations covering less than 50% of the water column. Simulations further show that the framework can also be extended to smoothly range-varying SSPs under a fixed flat-bottom condition. This study provides a physics-constrained framework for acoustic-field reconstruction and SSP inversion under sparse observations.