Approximating underwater sound propagation with neural networks: The case of normal mode theory

In shallow water environments and at low frequencies, normal mode theory provides an adequate description of the acoustic propagation; however, conventional normal mode simulation codes can be computationally intensive. To address this, we propose using neural networks (NNs) to approximate the modal wavenumbers and the modal depth functions. Predicting modal parameters using NNs is a relatively straightforward task compared to predicting more variable acoustic quantities, such as those involving modal interference like transmission loss. This approach allows NNs to be trained to approximate modal parameters across diverse environments and at different frequencies, thereby eliminating the need for expensive retraining of the NNs models. Once trained, the predicted modal wavenumbers and modal depth functions can be used according to the normal mode theory to obtain the acoustic field for any source and receiver positions. This approach divides the computation time to obtain the modal parameter by a factor of thirty compared to traditional normal mode codes, making it a promising solution for applications with limited computational resources, such as simulation on autonomous underwater vehicles.