SeaCast Delivers 15-Day Ocean Forecasts in 20 Seconds Using a Graph Neural Network

Ocean forecasting underpins maritime safety, aquaculture, coastal risk management and marine ecosystem monitoring. Traditional systems rely on physics‑based numerical models; for example, the Mediterranean Forecast System (MedFS) in CMEMS uses a bidirectionally coupled wave‑current model at ~4 km resolution to produce 10‑day forecasts, but requires 89 CPU cores and about 70 minutes per run.

Machine‑learning approaches have recently matched or exceeded global‑scale numerical forecasts, yet applying them to high‑resolution regional seas faces challenges such as irregular coastlines, complex lateral boundary conditions, and the need for fine vertical stratification.

To fill this gap, a joint team from the University of Helsinki, the Mediterranean Climate Change Research Center, and the University of Salento created SeaCast, a graph‑neural‑network (GNN) model specifically designed for regional ocean prediction. Key innovations include optimized graph construction to handle irregular marine grids, incorporation of near‑surface atmospheric forcing fields, and coupling of lateral boundary forcing to maintain consistency with global ocean circulation.

The study assembled a comprehensive dataset covering four categories: ocean state, atmospheric forcing, lateral boundary forcing, and satellite validation. Ocean state data stem from the Mediterranean Ocean Physical Analysis and Forecast System built on NEMO v4.2 and WAVEWATCH III v6.07 with 3‑D variational assimilation (OceanVar). Atmospheric forcing uses 2 m temperature, sea‑level pressure, and 10 m wind stress derived from ERA5 reanalysis, while boundary forcing draws on MedFS or global ocean forecasts for the Strait of Gibraltar and the Dardanelles. Satellite products (CMEMS L3S SST and sea‑level anomaly) provide validation data.

Training employed 35 years (1987‑2021) of daily reanalysis data for pre‑training (200 epochs) on 64 AMD MI250x GPUs (≈20.5 GPU‑hours) and 2 years (2022‑2023) of analysis data for fine‑tuning (30 epochs) on 8 GPUs (≈3.5 GPU‑hours). Validation used six months of 2024 analysis data (177 samples) and testing covered the remainder of 2024 plus early 2025, generating 15‑day forecasts from each daily initialization.

SeaCast follows an encoder‑processor‑decoder pipeline on a hierarchical graph grid adapted to the Mediterranean. The encoder maps ocean state and atmospheric fields to a coarse multi‑scale representation; the processor applies layered GNN operations to capture short‑ and long‑range interactions; the decoder reconstructs high‑resolution outputs. Unlike direct next‑step prediction, SeaCast learns daily change vectors, adds them to the current state, and incorporates dynamic boundary conditions in a self‑regressive loop, reducing bias from heterogeneous node neighborhoods compared with single‑scale GraphCast.

Performance-wise, SeaCast completes a full 15‑day forecast on a single GPU in 20 seconds, whereas MedFS needs 89 CPU cores and about 70 minutes for a 10‑day run. Across six variables (zonal/meridional flow, salinity, temperature, SST, sea‑level anomaly), SeaCast consistently outperforms MedFS, with the gap widening at longer lead times. Vertical analysis shows the greatest temperature and flow advantages near the surface, while salinity benefits appear deeper; performance converges around 192 m depth.

For extreme‑event detection, the authors defined marine heatwaves using the 90th‑percentile SST from satellite data. Both SeaCast and MedFS detect such events significantly better than a persistence baseline, with SeaCast achieving slightly higher Heidke Skill Scores, providing earlier warning within the 15‑day horizon.

Training‑period experiments reveal that using only 10 years of reanalysis data yields performance comparable to MedFS for flow, temperature and SST, while salinity and sea‑level anomaly require the full 35‑year record plus fine‑tuning to surpass MedFS. Fine‑tuning offers limited gains for sea‑level anomaly, likely due to sparse validation data.

The article concludes with a broader view of AI‑driven ocean forecasting: ECMWF’s AI Forecast System (AIFS) is entering operational use; NVIDIA’s Earth‑2 initiative integrates models like FourCastNet for global weather and climate; Google Research’s NeuralGCM combines differentiable dynamics with ML parameterizations. These efforts illustrate a shift from AI as an auxiliary tool toward core components of ocean prediction, promising improvements in accuracy, timeliness and interpretability.