Dr. Matjaž Ličer (NIB): Dynamical and deep learning approaches to ensemble storm surge modeling in the Adriatic basin
Northern Adriatic has been experiencing accelerated sea level rise since the 1990s, exacerbating the problem of regional storm-induced coastal floods. The importance of reliable and timely coastal flood forecasting has gained weight correspondingly but the this remains a challenging modeling problem. Adriatic Sea is a resonant, semi-enclosed elongated basin with well defined sea level eigenstates (seiches), which are typically excited by passing storms and which are close in period to those of semi-diurnal and diurnal tides, entering the basin from the Mediterranean. Total basin sea level is therefore highly sensitive to the phase lag between meteorologicaly induced seiches and gravitationaly induced tides. Consequently, even modest modeling errors in timing and trajectory of the storm often translate into substantial errors of sea level forecast. We can approach this problem probabilistically by directly solving Navier-Stokes and other governing equations in the Adriatic basin using an ensemble of realistic atmospheric boundary conditions. This is however numerically extremely expensive and compromises are often required on available HPC infrastructures. Deep learning architectures, used in computer vision, will be discussed as numerically cheap alternative with comparable forecasting skills. We will present an ensemble deep-neural-network-based sea level forecasting method HIDRA, which outperforms our set-up of the general ocean circulation model ensemble (NEMO v3.6) for all forecast lead times and at a minuscule fraction of the numerical cost. HIDRA exhibits larger bias but lower RMSE than our physics-based set-up over most of residual sea level bins. It introduces a trainable atmospheric spatial encoder and employs fusion of atmospheric and sea level features into a self-contained network which enables discriminative feature learning. We will further discuss on-going developments of HIDRA, among them its planned extensions to two-dimensional ocean surface modeling using linearized shallow-water wave equation over realistic bathymetry as the physics-informed part of the total loss function.
Žust, L., Fettich, A., Kristan, M., and Ličer, M.: HIDRA 1.0: deep-learning-based ensemble sea level forecasting in the northern Adriatic, Geosci. Model Dev., 14, 2057–2074 (2021), https://doi.org/10.5194/gmd-14-2057-2021.