Regional Ocean Forecasting with Hierarchical Graph Neural Networks

TL;DR

SeaCast is introduced, a neural network designed for high-resolution, medium-range ocean forecasting that employs a graph-based framework to effectively handle the complex geometry of ocean grids and integrates external forcing data tailored to the regional ocean context.

Abstract

Accurate ocean forecasting systems are vital for understanding marine dynamics, which play a crucial role in environmental management and climate adaptation strategies. Traditional numerical solvers, while effective, are computationally expensive and time-consuming. Recent advancements in machine learning have revolutionized weather forecasting, offering fast and energy-efficient alternatives. Building on these advancements, we introduce SeaCast, a neural network designed for high-resolution, medium-range ocean forecasting. SeaCast employs a graph-based framework to effectively handle the complex geometry of ocean grids and integrates external forcing data tailored to the regional ocean context. Our approach is validated through experiments at a high spatial resolution using the operational numerical model of the Mediterranean Sea provided by the Copernicus Marine Service, along with both numerical and data-driven atmospheric forcings.

Figures (35)

• Figure 1: Comparison of the SeaCast (AIFS forcing) 10-day lead temperature forecast at a depth of 22.7 m, initialized from simulation on August 1st, 2024, against the corresponding analysis field.
• Figure 2: Depth-averaged RMSE for a) temperature, b) salinity, and c) zonal velocity.
• Figure 3: Illustration of the Mediterranean Sea at all 18 depth levels. The surface layer is colored seagreen and the boundary forcing region in the Strait of Gibraltar is colored maroon. The color of the interior sea gets darker blue at increased depth. The visualized depth relative to the height and width is not to scale.
• Figure 4: Ocean variables are encoded onto a hierarchical mesh of the Mediterranean Sea shown here. Each layer has a different resolution allowing for interactions at different scales between observables.
• Figure 5: The top map displays satellite SST data from August 1st, 2024, with four labeled locations. The plots below show the evolution of the thetao_1 field at each location during a single forecast period, comparing the different models with satellite observation trends.