Physics-based and data-driven joint approach for low-frequency underwater acoustic field predictiona)

Area of Science:

• Ocean Acoustics
• Artificial Intelligence
• Naval Technology

Background:

• Accurate underwater acoustic field prediction is vital for naval operations.
• Traditional numerical models are computationally intensive, limiting real-time applications.
• Need for efficient and precise methods for low-frequency acoustic field prediction.

Purpose of the Study:

• Develop a hybrid physics-based and data-driven approach for efficient underwater acoustic field prediction.
• Improve prediction accuracy and computational speed compared to existing methods.
• Enhance capabilities for underwater target detection and autonomous vehicle navigation.

Main Methods:

• Utilized a convolutional autoencoder for bathymetric feature extraction.
• Employed a convolutional neural network to predict range-dependent modal coefficients.
• Integrated normal-mode theory and a residual network for acoustic field computation.

Main Results:

• Achieved 1.0-5.0 dB error improvement over adiabatic solutions in various environments.
• Outperformed end-to-end neural network baselines by up to 1.5 dB.
• Demonstrated a 180-200x computational speedup for modal coefficients calculation in deep-sea scenarios.

Conclusions:

• The hybrid AI approach offers significant improvements in accuracy and efficiency for low-frequency underwater acoustic field prediction.
• This method enhances performance, especially at 25 Hz, and is highly effective in deep-sea environments.
• The proposed model provides a computationally efficient solution for critical naval applications.