Physics-informed neural network with pretraining optimization for ocean acoustic field prediction
Juncong Tang, Haiqiang Niu
Abstract
To address the challenge of capturing high-frequency features in ocean acoustic field prediction using Physics-Informed Neural Networks (PINNs), this study introduces an enhanced pretraining optimization approach, termed PreT-OceanPINN, based on a recently proposed OceanPINN framework, a PINN-based method for predicting ocean acoustic pressure fields. By implementing a two-stage strategy-pretraining in a hypothetical environment followed by fine-tuning with real-world data-PreT-OceanPINN significantly improves both prediction accuracy for high-frequency components and training efficiency. In the pretraining stage, the model is trained using envelope signals derived from a simulated environment, allowing it to internalize the underlying physical principles of sound propagation. During the fine-tuning stage, a limited amount of measured data are used to adapt the model to complex real-world conditions. Compared to the standard OceanPINN approach, PreT-OceanPINN delivers more accurate high-frequency predictions without increasing real data dependency, thus demonstrating clear performance advantages. The effectiveness of the proposed method is validated through numerical simulations and experimental data from the SWellEx-96 field experiment.