Physics-informed neural networks to reconstruct surface velocity field from drifter data

Accessing ocean velocity data is critical to improving our understanding of ocean dynamics, which affects our prediction capabilities for a range of services that the ocean provides. Because ocean current velocity information is scarce, prediction efforts have mostly relied on numerical models of ocean physics to reconstruct and predict velocity fields at desired spatial and temporal resolutions. However, numerical models, by design, are a simplified representation of the physics laws that govern ocean dynamics, hence error-prone even with data assimilation. Although accurate measurements of the flow field can be obtained using ocean drifters along their trajectories, their Lagrangian nature and sparsity make them unfit to provide direct Eulerian measurements. To address this issue, we apply a deep learning model called Physics-Informed Neural Networks (PINN) to reconstruct ocean surface velocity fields using sparse measurements obtained from drifters. We show that the physics learning part of the network is essential for the accurate reconstruction of the velocity field. In particular, we show the poor performance of the same deep neural network without the physics part, which reveals the ability of the partial differential equations derived by the PINN to capture the flow features’ dynamics. Our method is evaluated on both virtual and real drifters datasets. The reconstructed flow fields reveal the role of small-scale features in improving the representation of mesoscale flow dynamics.