SeaCast Generates 15-Day Mediterranean Forecasts in 20 Seconds on a Single GPU

Physical ocean models are computationally expensive. Running the equations that govern fluid dynamics across a regional domain at high spatial resolution requires dozens to hundreds of CPUs working in parallel for minutes to hours. That computational cost limits how many scenarios can be tested, how quickly forecasts can be updated, and how feasible it is to run probabilistic ensembles - multiple simulations that collectively map out forecast uncertainty.

SeaCast, a new AI-based forecasting system for the Mediterranean Sea developed jointly by the Euro-Mediterranean Center on Climate Change (CMCC) and the University of Helsinki, changes that cost equation substantially. The model produces a 15-day high-resolution forecast in approximately 20 seconds on a single GPU. The equivalent Copernicus operational physics-based model takes around 70 minutes running on 89 CPUs to produce a 10-day forecast - a shorter prediction horizon completed in roughly 200 times as much compute time.

What SeaCast does differently

Existing global AI ocean models operate at coarser spatial resolutions and typically rely solely on ocean state variables as inputs. SeaCast addresses both limitations. It operates at approximately 4 km resolution - 1/24 degree - matching the CMCC Mediterranean Forecasting System (MedFS) delivered through the Copernicus Marine Service. And it integrates atmospheric variables alongside ocean state variables during both training and forecasting, not just at the surface boundary.

That atmospheric integration matters because the Mediterranean surface is strongly coupled to the atmosphere - wind stress, heat flux, and freshwater input from precipitation all drive surface currents and mixed-layer dynamics. Models that treat the atmosphere as a simple boundary condition miss the variability in those forcings. SeaCast's sensitivity experiments identified which atmospheric variables contribute most to improved predictions, and the results confirmed that including them substantially improves forecast accuracy near the surface.

The architecture uses a graph-based neural network rather than a conventional grid-based approach. Graph networks handle irregular geometries more naturally than structured grids, which is relevant for the Mediterranean's complex coastlines, islands, and semi-enclosed sub-basins. The model also accounts for lateral boundary conditions - a technical challenge in regional models where open-ocean boundaries need to be specified consistently.

Performance benchmarks

SeaCast consistently outperforms the Copernicus operational model over the standard 10-day forecast horizon and extends skillful predictions to 15 days. The model was trained on CMCC Mediterranean reanalysis data covering up to 35 years of historical ocean states, available through the Copernicus Marine website. Longer training periods - up to the full 35-year dataset - improved model skill, suggesting the algorithm benefits from capturing decadal-scale variability in the training distribution.

Forecasts extend to a depth of 200 meters, covering the upper ocean layer where most shipping, ecological, and coastal management applications are concentrated.

Speed as a scientific tool, not just efficiency

The 20-second runtime enables something the operational model cannot practically do: rapid scenario testing and probabilistic ensemble generation. Running dozens or hundreds of model variants to map out forecast uncertainty requires either a very large compute cluster or very fast individual runs. With SeaCast, a 50-member ensemble can be completed on a single GPU in well under an hour.

"This achievement demonstrates how bringing together oceanography, atmospheric science, and AI expertise produces tangible results, and can unlock a new generation of regional ocean forecasts," said Emanuela Clementi, CMCC researcher and co-author of the study. "Combining physical insight with advanced AI allows us to improve forecast accuracy while dramatically reducing computational costs."

Practical applications and what comes next

Mediterranean Sea forecasting serves commercial shipping, aquaculture operations, environmental monitoring, and coastal hazard management - from early warning for storm surges to tracking pollution dispersion. Faster and more reliable forecasts translate to better operational decisions in each of those domains.

CMCC researchers are now working to integrate SeaCast into operational forecasting chains alongside traditional physics-based models. Hybrid approaches, where AI models supplement or correct physics-based outputs, have shown promise in other domains. Whether SeaCast eventually replaces or consistently outperforms the physics-based operational system across all conditions and seasons requires further validation.

The collaboration between CMCC in Italy and the University of Helsinki produced a model trained on freely available public data - the full reanalysis dataset is accessible through the Copernicus Marine website, which means other research groups can reproduce and extend this work.