SeaCast is a high-resolution AI forecasting system developed for the Mediterranean that aims to deliver faster and more energy-efficient ocean predictions than conventional numerical models. The system is designed for regional complexity rather than broad global coverage, using an architecture that can represent irregular coastlines and boundary effects that typically make regional forecasting harder than open-ocean prediction. By targeting a basin with dense coastal activity and fast-changing conditions, the project positions forecasting speed and accuracy as operational advantages rather than academic benchmarks.

 

Model Design Built for Regional Geometry and Boundaries

 

The system performs autoregressive forecasting using a graph neural network, which allows the model to operate on mesh representations instead of being constrained to simple rectangular grids. In practical terms, this lets SeaCast encode sea-state and atmospheric inputs onto a coarser mesh, process them through graph layers designed to capture spatial relationships, and then decode predictions back onto the original grid. Boundary conditions are explicitly incorporated so forecasts remain physically consistent at lateral edges, addressing a common weakness in regional AI models that can drift when they are forced by changing inflows and outflows.

 

Resolution, Depth and Training Data

 

SeaCast operates at roughly 4 km resolution, aligned with the resolution of the CMCC Mediterranean operational forecasting system MedFS distributed through Copernicus Marine Service. It generates forecasts down to 200 metres depth, and it was trained on CMCC Mediterranean reanalysis data provided at the same resolution through Copernicus Marine. The alignment between training data and operational resolution is important because it reduces the need for aggressive downscaling and helps the model learn basin-specific dynamics at the scale relevant for coastal and shelf processes.

 

Read more: Hybrid Offshore Platforms Could Lift Output by 70 Percent While Cutting Costs, University of Surrey Review Finds

 

Performance Against the Operational Baseline

 

In reported results, SeaCast consistently outperforms the Copernicus operational system over the standard 10-day forecast horizon and extends useful predictions out to 15 days. The most dramatic difference is computational efficiency. While the operational numerical system takes roughly 70 minutes using 89 CPUs to produce a 10-day forecast, SeaCast is described as generating a 15-day forecast in about 20 seconds on a single GPU. This shift changes the practical economics of forecasting by making frequent updates, rapid iteration, and larger ensembles feasible without a proportional increase in compute budgets.

 

Why Adding Atmospheric Inputs Improves Skill

 

A key technical point is that SeaCast integrates atmospheric forcing data along with ocean variables during training and forecasting, rather than relying primarily on ocean-only signals. Results described in the study indicate that atmospheric information meaningfully improves accuracy, especially near the surface where winds, heat fluxes, and pressure patterns strongly shape currents, temperature, and short-term variability. Sensitivity testing is used to identify which atmospheric variables contribute most and to show how longer training histories, up to 35 years of data, can strengthen model skill by exposing it to a wider range of regimes and extremes.

 

Operational Use Cases and the Next Step

 

Faster forecasting has direct operational value because it enables what-if scenario testing and probabilistic ensemble forecasting at a cadence that is difficult to achieve with purely physics-based systems. For the Mediterranean, where shipping, aquaculture, coastal monitoring, and hazard planning depend on timely and localised information, the practical benefit is less time spent waiting for model outputs and more capacity to quantify uncertainty. The next stage described by the researchers is integrating SeaCast into operational forecasting chains alongside traditional models, using AI to complement physics-based approaches rather than replacing them outright, with the goal of improving decision support while cutting computational cost.