Physics-informed neural networks for predicting ocean spatio-temporal fields

Simulations of the ocean environment are important for assessing acoustic variability due to large and small scale changes in the propagation medium. By incorporating the Internal Gravity Waves (IGWs) partial differential equation (PDE) in ocean models, such as the HYbrid Coordinate Ocean Model (HYCOM) system, it is possible to model the impact of tides on acoustic propagation. This requires HYCOM simulations of the ocean environment, which are computationally expensive, especially at the high vertical resolutions required to capture IGWs. In this work, we develop a machine-learning approach to predicting the effects of IGWs without incurring the computational cost of high-resolution numerical modeling. We apply a deep-learning approach in the form of a Physics Informed Neural Network (PINN) that incorporates the IGW PDE system to make predictions about oceanographic quantities of interest with internal tides at high vertical resolution. Training data comprise HYCOM 4D fields (time, depth, latitude, and longitude) both with (IGW-HYCOM) and without tides in the Pacific subtropical/equatorial region. The developed framework is able to map a relationship between lower-resolution HYCOM inputs to highly resolved ocean fields that include internal tides.