Efficient and high-precision underwater acoustic field prediction is crucial for underwater target detection, autonomous vehicle path planning, and other naval applications. Traditional numerical models suffer from high computational complexity. This study proposes a hybrid physics-based and data-driven approach for low-frequency underwater acoustic field prediction. First, the convolutional autoencoder is constructed to extract bathymetric features. These, along with source depth, are input into a convolutional neural network to predict range-dependent modal coefficients, which are integrated with normal-mode theory to compute the acoustic field. Finally, the residual network further refines the prediction. Using coupled-mode solutions as ground truth, the proposed neural network achieves error improvement margins of 1.0-3.0 dB in shallow-water environments, and 1.0-5.0 dB in deep-sea conditions, compared to adiabatic solutions. Relative to end-to-end neural network baselines, the present method delivers performance improvements of approximately 1.5 dB in shallow-water scenarios, and approximately 0.7 dB in deep-sea environments, with particularly enhanced performance at 25 Hz. For low-frequency acoustic field computation in deep-sea settings, the neural network demonstrates a 180-200 times computational speedup in modal coefficients calculation over numerical models, significantly enhancing the efficiency of acoustic field prediction.
Feng et al. (Wed,) studied this question.