Machine Learning for the complex response of the Wadden Sea

The currents, the sea level, and the transport of material (pollutants and nutrients, for example) in the Wadden Sea depend in a complex way on the forcing: the tides, the wind, the atmospheric pressure, and the fresh water discharged from Lake IJssel. It is possible to model these complex interactions using simulations, but these simulations are expensive. They usually have to be run hundreds of computing cores over several days. This is problematic, for example, for predictions of the future of the Wadden Sea
under several distinct climate change scenarios. In this project, we look at the possibilities of using Machine Learning to predict certain quantities that characterize the hydrodynamic response of the Wadden Sea as a function of the forcing conditions. We use a particular type of Recurrent Neural Networks (RNN) called long-short-term memory (LSTM) RNN. The advantage of this type of RNN is that it can process sequences of data. This is crucial because the state of the Wadden Sea at any given time depends not only on the
forcing at that time but also on the state and the forcing in the immediate past. To train and test the network, we use data from a simulation spanning 36 years into the past. In general, the RNN LSTM models can recover the most important features. Further research needs to be done still to improve the ML to be use for future prediction.