Efficient deep learning surrogate method for predicting the transport of particle patches in coastal environments

Abstract

Several coastal regions require operational forecast systems for predicting the transport of pollutants released during marine accidents. In response to this need, surrogate models offer cost-effective solutions. Here, we propose a surrogate modeling method for predicting the residual transport of particle patches in coastal environments. These patches are collections of passive particles equivalent to Eulerian tracers but can be extended to other particulates. By only using relevant forcing, we train a deep learning model (DLM) to predict the displacement (advection) and spread (dispersion) of particle patches after one tidal period. These quantities are then coupled into a simplified Lagrangian model to obtain predictions for larger times. Predictions with our methodology, successfully applied in the Dutch Wadden Sea, are fast. The trained DLM provides predictions in a few seconds, and our simplified Lagrangian model is one to two orders of magnitude faster than a traditional Lagrangian model fed with currents.