Physics Informed Neural Networks for Modeling Large-Scale Wind Driven Ocean Circulation

Abstract

This study investigates the application of the physics informed neural network as a meshfree collocation method for approximating solutions to large-scale wind driven ocean circulation models. By integrating the Stommel and Stommel-Munk models into the neural network framework, the neural network provides a viable alternative to traditional numerical methods for simulating ocean circulation. The architecture of the neural network was systematically optimized through hyperparameter tuning, including the selection of optimizers, activation functions, network configurations, and learning rate schedulers to ensure stable convergence and minimize fluctuations in training loss. The effects of different training point distributions, such as uniform, uniform-refined, random, and random-refined, were also examined. The results show that refining the distribution of training points near the western boundary layer can achieve similar accuracy and training performance even with fewer points. This approach highlights the potential of the physics informed neural network to address more complex oceanographic models, where conventional numerical methods may be constrained by data availability and computational cost.

Keywords

• Deep learning
• ocean currents
• physics informed neural network
• quasi-geostrophic
• Stommel
• Stommel-Munk