Predicting oceanic Lagrangian trajectories with hybrid space-time CNN architecture

Abstract. Lagrangian dynamics simulation is a challenging task, as it typically depends on integrating velocity fields, whose estimation is inherently difficult due to both theoretical and technical constraints. Neural Network approaches provide practical methods to overcome most of related complications by learning directly from data. In this paper a deep Convolutional Neural Network (CNN) for Lagrangian trajectories simulation is presented. The proposed architecture is inspired by existing Computer Vision methods, combining Long-Short Term Memory and U-Net architectures to enforce causality. Several training setups are considered, including conditional Generative Adversarial Network (cGAN) training. The results are evaluated using Lagrangian metrics.