Researchers have developed an artificial intelligence–driven ocean forecasting system named SeaCast that can generate highly detailed 15-day predictions of Mediterranean Sea conditions in just seconds — a dramatic improvement in speed and resolution over traditional numerical models. The model, described in a paper published in the journal Scientific Reports, was developed by scientists at the Euro-Mediterranean Center on Climate Change (CMCC) in collaboration with the University of Helsinki and uses a graph neural network (GNN) to capture complex ocean and atmospheric interactions.
Unlike many existing global AI forecasting models that work at coarse resolutions relying mainly on ocean data, SeaCast integrates both oceanic and atmospheric variables. This allows it to better represent the intricate regional dynamics of the Mediterranean, including detailed coastlines, lateral boundaries, and interactions between the sea surface and atmosphere. The system operates at about 4 km horizontal resolution (1/24°) and can produce forecasts down to a depth of 200 m.
SeaCast significantly outperforms traditional operational models, which typically require computing clusters and tens of minutes to hours to generate forecasts. The Mediterranean Forecasting System (MedFS), for example, usually takes around 70 minutes on 89 central processing units (CPUs) to produce a 10-day forecast. In contrast, SeaCast can deliver a full 15-day forecast in approximately 20 seconds using a single graphics processing unit (GPU). This efficiency opens the door to rapid scenario testing, probabilistic ensemble forecasting, and extensive uncertainty analysis — tools that are valuable for both research and practical applications.
High-resolution ocean forecasts are vital for a wide array of sectors including shipping, aquaculture, environmental monitoring, and coastal risk management. Fast and accurate predictions can assist authorities and stakeholders in proactive planning and responses to changing sea conditions. SeaCast’s ability to combine physical insight with advanced AI methods represents a significant leap forward in regional ocean forecasting.
One of SeaCast’s key innovations is the integration of atmospheric forcing data during both model training and forecasting, which substantially improves accuracy, especially near the sea surface where atmospheric effects are strongest. Sensitivity experiments in the research show the relative contribution of different atmospheric variables to forecast skill and highlight the importance of long training periods — up to decades of historical data — for achieving the best performance.
Looking ahead, the researchers plan to integrate SeaCast into operational forecasting systems, complementing traditional physics-based approaches. As the first high-resolution, regional AI ocean model, SeaCast is setting a new benchmark for ocean prediction and offers a path toward faster, smarter, and more reliable marine forecasts worldwide.
