Neural Prediction of Lagrangian Drift Trajectories on the Sea Surface

Abstract This study proposes a new deep learning approach for the simulation of Lagrangian drift at the sea surface with the objective to overcome the current limitations of the existing model-based and learning-based methods. The proposed framework, called DriftNet, is inspired by the Eulerian Fokker–Planck representation of Lagrangian drift. DriftNet is able to simulate the Lagrangian trajectory of a fluid parcel given the corresponding Eulerian sea surface currents and the spatially explicit encoding of the parcel’s initial position. The efficacy of DriftNet is demonstrated through three benchmarks: two benchmarks involving fully simulated ocean data and one combining operational ocean reanalysis model along with in situ drifters’ data. The study focuses on two regions with different dynamical regimes: the northeast Pacific and the Gulf Stream. The findings indicate that DriftNet outperforms current state-of-the-art model-based and learning-based methods. Additionally, this study explores how the sea surface height—derived from both modeled and observed ocean—affects Lagrangian drift simulations when used as input to DriftNet. Significance Statement We propose a novel deep learning model, DriftNet, for the conditional generation of Lagrangian trajectories on the sea surface. Our model is based on Eulerian Fokker–Planck formalism and can be conditioned by multiple geophysical fields. We highlight the overall overperformance of DriftNet compared to the baseline model-based and learning-based approaches. We put in evidence the capacity of the proposed method to extract pertinent information from various geophysical fields, from both modeled and observed data. We highlight the significant impact of the observed sea surface height when combined with sea surface currents in the quality of the generation of Lagrangian trajectories.