Forecast of Particle Spreading Using Machine Learning in a Complex Multiple-Inlet Coastal System

The implementation of continuous operational forecast systems using numerical models for coastal environments are scarce, computationally expensive, and difficult to maintain. As an alternative, computationally cheaper tools such as machine learning models can be employed. This is especially relevant when the time to produce a forecast is paramount like in oil spills, marine litter spread due to container-ship accidents, and search and rescue operations. Working in this direction, we tested the skill of an advanced deep learning model, namely a convolutional long short-term memory network (ConvLSTM), to predict the Lagrangian advection (the displacement vector of the center of mass) and the dispersion (the spread described by a covariance matrix) of patches of passive tracers. This model was trained with data from a realistic numerical simulation of the Dutch Wadden Sea: a multiple-inlet system of great ecological importance. Using the relevant drivers (wind, tidal amplitude, and atmospheric pressure), the model was set to learn the advection and dispersion after one tidal period of clouds of particles released on a 200 x 200 m grid, covering the entire DWS. Our results show that the model learned the system-wide temporal variability of both advection and dispersion, while the local spatial features were better reproduced for advection than for dispersion. We use the predicted advection and dispersion as inputs to a set of stochastic differential equations for the reconstruction of particle trajectories, as it is commonly done in particle tracking applications that employ diffusion instead of dispersion. We were able to predict the temporal evolution over several tidal periods of particle patches released from specific locations under contrasting cases like calm and stormy conditions. Our method was employed to predict only the horizontal spreading, but it can be extended to predict the 3D evolution of the particle clouds. Finally, our approach requires simulation data and relevant drivers (e.g. atmospheric forcing and tidal amplitudes) for training and the same drivers from any typical forecast systems for forecasting the evolution of particle patches, which makes it a promising operational tool.