TL;DR
HaiRuo is a 1.05-billion-parameter foundation model that revolutionizes ocean eddy forecasting by treating eddies as both continuous physical fields and discrete moving objects simultaneously. By pretraining on 24 years of satellite data and employing a novel Lagrangian decoder, it reduces 15-day eddy trajectory errors by 18% globally and maintains physical consistency without explicit hard-coded laws.
What Problem Does This Paper Address?
Mesoscale eddies—massive swirling vortices 50–300 km across—are the "weather" of the ocean. they carry the bulk of the ocean's kinetic energy and are responsible for transporting heat, salt, and nutrients across basins. Despite their importance, forecasting where an eddy will be two weeks from now has traditionally been a fragmented process.
Standard AI ocean models forecast "fields" (like sea-surface height) and then rely on separate, heuristic tracking algorithms to identify eddies after the fact. This "post-hoc" approach is structurally flawed; the tracker doesn't understand the model's internal dynamics, leading to trajectories that often drift far from reality. The authors argue that the "object" (the eddy) belongs inside the decoder, allowing the model to re-sample its own evolving forecast as it predicts the path.
Methods
The HaiRuo system consists of two primary components with a clear division of labor:
- HaiRuo-1B (The Backbone): A massive foundation model pretrained via self-supervision on 24 years of unlabeled global satellite altimetry. It uses a two-stage spherical stem to process data without aliasing and a "dual-rate" memory system to handle a 30-day historical context efficiently. It is trained on three objectives: future prediction, masked reconstruction, and a "flow-aligned contrast" that pairs tokens based on geostrophic velocity.
- XuanJi (The Decoder): A specialized Lagrangian object decoder. Instead of looking at static maps, XuanJi uses "object attention" to sample the model’s own 15-day forecast in a co-moving polar frame. It iteratively refines the trajectory across three "look-ahead" rounds, essentially "feeling" the predicted environment to adjust the eddy's path.
Datasets
- Pretraining: 24 years (1993–2016) of global satellite altimetry maps (DUACS), totaling over 8,700 daily windows.
- Fine-tuning Labels: The META4.0 global mesoscale eddy atlas, which provides "ground truth" for eddy masks and trajectories.
- Evaluation: An independent global test set from 2020–2022 (Test48) and a development set from 2017–2019 (Dev48). The evaluation is truly global, covering regions never seen during the labeled training phase.
Key Insights and Ideas
- The "Object" in the Decoder: The most significant conceptual shift is moving eddy tracking from a post-processing step into the model's architecture. By sampling azimuthal Fourier modes along a predicted orbit, the model captures the physical asymmetries (dipoles and quadrupoles) that drive an eddy's self-advection and strain.
- Foundation Models Lift All Boats: Pretraining on decades of unlabeled data provides a "physical prior" that improves every downstream task. Even with a frozen backbone, HaiRuo-1B outperforms specialized models trained from scratch.
- Emergent Physics: Without being programmed with the laws of fluid dynamics, HaiRuo reproduces complex phenomena like westward drift, the eastward reversal at high southern latitudes, and the tendency for cyclones to drift poleward relative to anticyclones.
Key Results
HaiRuo was tested against 14 baseline methods and specialized global ocean forecasters like WenHai and XiHe.
- Trajectory Accuracy: It achieved a 15-day center error of 25.20 km, an 18% improvement over the strongest baseline (PredRNN-V2 at 30.58 km).
- Field Skill: It significantly improved the accuracy of sea-level anomaly (SLA) forecasts, reaching an RMSE of 4.65 cm at day 15.
- Robustness: The model maintained its lead across different phases of the El Niño–Southern Oscillation (ENSO) and in "out-of-distribution" geographic regions not used for labeled training.
- Pipeline Efficiency: When compared to existing operational forecasters, HaiRuo's direct object trajectories cut pipeline error by 77–78% globally.
Limitations and Future Work
- Data Latency: The study uses "delayed-time" satellite maps for high accuracy. Transitioning to operational, real-time forecasting would require adapting the model to "near-real-time" data streams which are typically noisier.
- Dimensionality: Currently, the model focuses on 2D surface dynamics. Extending this foundation model approach to 3D ocean volumes (incorporating depth profiles) is a natural next step for more comprehensive oceanographic modeling.
- Computational Requirements: While the authors successfully trained the 1B parameter model on a single GPU in about 12 hours, the inference requirements for a global, high-resolution operational system would be substantial.
Summary updated