Physics-informed neural network-based predictions of ocean acoustic pressure fields

Abstract

Physics-informed neural network (PINN) trains the network using sampled data and encodes the underlying physical laws governing the dataset, such as partial differential equations (PDEs). A trained PINN can predict data at locations beyond the sampled data positions. The ocean acoustic pressure field satisfies PDEs, Helmholtz equations. We present a method utilizing PINN for predicting the underwater acoustic pressure field. Our approach trains the network by fitting sampled data, embedding PDEs, and enforcing pressure-release surface boundary conditions. We demonstrate our approach under various scenarios. By incorporating PDE information into a neural network, our method captures more accurate solutions than purely data-driven methods. This approach helps enhance the information content of sampled data when dealing with a limited amount of data.