Research map / Time series / Long-term forecasting
FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces FuXi-Ocean, a data-driven global ocean forecasting model producing six-hourly predictions at 1/12° resolution down to 1500 m, using a Mixture-of-Time module that selects among predictions from multiple temporal input windows.
Problem
Data-driven ocean forecasting has been limited to daily resolution and shallow depths, and autoregressive models accumulate errors when forecasting at sub-daily intervals with diverse variable-specific temporal dynamics.
Approach
A Mixture-of-Time (MoT) module that keeps a per-channel reliability matrix updated from spatially averaged MAE of predictions made from different historical windows, and uses the top-K most reliable windows per variable/depth to form the final forecast, plus a context-aware (time/space-modulated) encoder and AdaLN-based attention predictor.
Key results
- On 2022 IV-TT SST comparison, FuXi-Ocean's RMSE is reported as lower than HYCOM, BLK, and FOAM, with significantly smaller cumulative error growth (6-hourly outputs averaged to daily means).
- Removing the MoT module increases RMSE for salinity, temperature, and SSH by up to nearly 40% across lead times on the HYCOM-RD 2015 test set.
- Trained on about 8.5 years of HYCOM-RD data (2006-2014), versus longer 1993-2020 ranges used by Xihe, Langya and WenHai.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FuXi-Ocean instead of operational forecast systems | FuXi-Ocean · IV-TT 2022 SST | RMSE | Helped |
| add the Mixture-of-Time module (ablation) | FuXi-Ocean · HYCOM-RD 2015 salinity, temperature, and SSH | RMSE: nearly 40% | Helped |
| add the Mixture-of-Time module (ablation) | FuXi-Ocean · HYCOM-RD 2015 ocean currents U and V | RMSE | No clear effect |
| use four rather than two input time steps (ablation) | FuXi-Ocean without MoT · HYCOM-RD 2015 temperature and salinity beyond 48 hours | RMSE | Helped |
| use four rather than two input time steps (ablation) | FuXi-Ocean without MoT · HYCOM-RD 2015 ocean currents U and V | RMSE | No clear effect |
| produce six-hour forecasts rather than daily forecasts | FuXi-Ocean · HYCOM-RD 2015, all variables | RMSE | Helped |
Surprising findings
- Ocean current U/V forecasts are much less sensitive to removing MoT or reducing history from four to two steps than T, S and SSH.
- Raw Argo/IV-TT salinity and temperature observations had errors too large to use directly, so evaluation was restricted to SST.
Limitations
- Trained only on reanalysis data (HYCOM-RD), inheriting biases of the underlying numerical model.
- Evaluation relies mainly on RMSE and SST against observations; salinity and temperature observation validation was unreliable; no phenomenon-specific metrics.
- Long-term (seasonal to interannual) forecast skill unexplored; K=1 choice may not hold for larger history windows.
Experimental setup
Datasets: HYCOM-RD reanalysis GODAE OceanView IV-TT Class 4 observations Argo profiles CMEMS buoy data
Benchmarks: HYCOM-RD 2015 test set GODAE OceanView IV-TT Class 4 (2022 SST) CMEMS buoy currents
Metrics: latitude-weighted RMSE
Baselines: HYCOM BLK FOAM WenHai GLORY12 woMoT woMoT(2times)
Builds on: FuXi Swin-Transformer AdaLN Charbonnier loss
Closest papers by meaning
- OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting (NeurIPS 2026 · similarity 0.86)
- Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting (NeurIPS 2026 · similarity 0.86)
- OxyGenerator: Reconstructing Global Ocean Deoxygenation Over a Century with Deep Learning (ICML 2024 · similarity 0.85)
- MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting (CVPR 2025 · similarity 0.84)
- CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes (ICML 2025 · similarity 0.84)
- IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.