Physics-informed neural network with pretraining optimization for ocean acoustic field predictiona).

Area of Science:

• Ocean acoustics
• Computational physics
• Machine learning

Background:

• Physics-Informed Neural Networks (PINNs) are used for ocean acoustic field prediction.
• Capturing high-frequency features in ocean acoustics remains a challenge.
• The OceanPINN framework offers a PINN-based solution for predicting ocean acoustic pressure fields.

Purpose of the Study:

• To introduce an enhanced pretraining optimization approach, PreT-OceanPINN, for improved ocean acoustic field prediction.
• To enhance the accuracy of high-frequency component prediction and training efficiency.
• To validate the effectiveness of PreT-OceanPINN using numerical simulations and experimental data.

Main Methods:

• Implemented a two-stage strategy: pretraining in a simulated environment and fine-tuning with real-world data.
• Pretraining involves training the model using envelope signals from a simulated environment to learn physical principles of sound propagation.
• Fine-tuning utilizes limited measured data to adapt the model to complex real-world conditions.

Main Results:

• PreT-OceanPINN significantly improves prediction accuracy for high-frequency acoustic components.
• The enhanced approach demonstrates improved training efficiency compared to standard OceanPINN.
• PreT-OceanPINN achieves higher accuracy without increased dependency on real-world data.

Conclusions:

• PreT-OceanPINN offers a superior method for ocean acoustic field prediction, particularly for high-frequency features.
• The two-stage pretraining and fine-tuning strategy effectively enhances model performance.
• The method shows clear performance advantages validated by numerical and experimental results from the SWellEx-96 experiment.