Daria Botvynko

Deep Learning for Lagrangian Drift Simulation on Sea Surface

23-Apr-24

The simulation of Lagrangian trajectories on the ocean surface holds significance across various application domains, from monitoring plastic and debris movement to investigating algae and plankton dynamics, and forecasting trajectories crucial for search and rescue operations. Assessing the capabilities of ocean numerical models in accurately representing small-scale dynamics is also vital. However, generating realistic trajectories on the sea surface poses a notable scientific challenge within operational oceanography. Model-based methods rely on advection procedures using sea surface velocity fields, yet discrepancies in these fields can lead to inaccurate trajectory modeling. Data-driven learning-based methods have shown promise in capturing spatio-temporal dependencies in simulated trajectories, but few have been applied to the conditional simulation of individual Lagrangian trajectories. Addressing these limitations, this study introduces DriftNet, a novel Deep Learning framework for the conditional simulation of individual trajectories on the sea surface. DriftNet can be trained on any geophysical field containing ocean dynamics information and generates trajectories based on a spatially-explicit latent encoding of targeted trajectory, inspired by Eulerian Fokker-Planck formalism. This approach allows for non-local feature extraction from conditioning input fields, ensuring that the entire dynamics of the surrounding area are considered in modeling the simulated trajectory.