Exploring ML-based decadal predictions of the German Bight storm surge climate

Storm surges and elevated water levels regularly challenge coastal protection and inland water management along the low-lying coastline of the German Bight. Skillful seasonal-to-decadal (S2D) predictions of the local storm surge climate would be beneficial to stakeholders and decision makers in the region. While storm activity has recently been shown to be skillfully predictable on a decadal timescale with a global earth system model, surge modelling usually requires very fine spatial and temporal resolutions that are not yet present in current earth system models. We therefore propose an alternative approach to generating S2D predictions of the storm surge climate by training a neural network on observed water levels and large-scale atmospheric patterns, and apply the neural network to the available model output of a S2D prediction system. We show that the neural-network-based translation from large-scale atmospheric fields to local water levels at the coast works sufficiently well, and that several windows of predictability for the German Bight surge climate emerge on the S2D scale.