Computer Science & AI
AI Model MedFormer Outperforms Traditional Ocean Forecasts in Mediterranean Sea
Original Authors: Epicoco, Donno, Accarino, Norberti, Grandi, McAdam, Elia, Clementi, Nassisi, Scoccimarro, Coppini, Gualdi, Aloisio, Masina, Boccaletti, Navarra
Accurate ocean forecasting is crucial for many marine activities, yet extending successful AI methods from atmospheric forecasting to complex ocean systems has been difficult due to their slower dynamics and intricate boundary conditions. This research introduces MedFormer, a cutting-edge, fully data-driven deep learning model specifically engineered for medium-range ocean forecasting within the Mediterranean Sea. Its development addresses the growing need for more precise and efficient predictive tools in oceanography.
MedFormer is built on a U-Net architecture, enhanced with 3D attention mechanisms, allowing it to operate at a high horizontal resolution of 1/24°. The model is trained on 20 years of daily ocean reanalysis data, further refined with high-resolution operational analyses. Employing an autoregressive strategy, MedFormer produces 9-day forecasts, leveraging both historical ocean states and atmospheric forcings to make it well-suited for operational use.
The efficacy of MedFormer was rigorously tested against the Mediterranean Forecasting System (MedFS), a state-of-the-art traditional numerical model from CMCC. Using both analysis data and independent observations, MedFormer consistently demonstrated superior forecast skills across key 3D ocean variables. Metrics like Root Mean Squared Difference and Anomaly Correlation Coefficient confirmed its enhanced accuracy. As lead author Epicoco notes in the paper, "These findings underscore the potential of data-driven approaches like MedFormer to complement, or even surpass, traditional numerical ocean forecasting systems in both accuracy and computational efficiency."