TL;DR
This paper introduces TianHai, a novel data-driven global ocean forecasting model capable of 6-hour predictions at high spatial resolution. It achieves state-of-the-art accuracy for key ocean variables, outperforming traditional numerical models and existing deep learning approaches, while significantly reducing computational costs.
What Problem Does This Paper Address?
Accurate, high-resolution ocean forecasting is crucial for everything from predicting natural hazards to optimizing maritime operations. While traditional numerical ocean models can achieve this, they are computationally expensive and often struggle with maintaining predictive skill at fine scales. Existing data-driven models offer computational efficiency but are typically limited to daily forecasts, failing to capture crucial sub-daily oceanic phenomena like diurnal temperature variations and mesoscale eddies. This paper tackles the challenge of developing a data-driven model that can deliver both high-resolution (eddy-resolving) and high-frequency (sub-daily) global ocean forecasts.
Methods
TianHai employs a sophisticated autoregressive deep learning architecture designed to capture multi-scale temporal dynamics. Key components include:
- Feature Extraction Module: A shared encoder processes input data, modulated by a "prior information network" (M) that incorporates spatiotemporal context (e.g., diurnal patterns, coordinates). This allows for efficient representation of global ocean data.
- Prediction Module: This module utilizes stacked attention blocks with adaptive layer normalization (AdaLN) and feed-forward networks to model the complex, non-linear evolution of oceanic states.
- MoT Module: A novel "multi-timescale temporal" module adaptively integrates predictions across different temporal scales, addressing the limitation of fixed temporal resolutions in previous models.
- Hybrid Approach: TianHai integrates near-surface atmospheric forcings (wind) directly, creating a fully data-driven framework for coupled air-sea prediction without relying on numerical atmospheric models.
- Training Strategy: A two-stage training approach (single-step prediction followed by multi-step fine-tuning) is used to improve forecast stability. A latitude-weighted loss function is employed to account for the varying grid cell areas at different latitudes.
- Depth-Dependent Land-Sea Masking: Unlike some models that use a uniform mask, TianHai applies depth-dependent masks to ensure physically consistent predictions across all vertical layers.
Datasets
- Training/Validation/Testing: HYCOM Global 1/12° Reanalysis and Analysis datasets (January 2006 - December 2021).
- Evaluation:
- HYCOM Global 1/12° analysis data (2022) for vertical profile evaluation.
- Argo float observations (quality-controlled near-real-time data) for temperature and salinity profiles.
- AVISO+ altimetry data for sea level anomaly (SLA).
- Global Drifter Program (GDP) data for SST and 15-m zonal/meridional currents.
- The GODAE Ocean View IV-TT Class 4 framework, which provides interpolated observational data for direct comparison.
- Typhoon Aere (2022) SST data for extreme weather event testing.
Key Insights and Ideas
- Sub-daily is Crucial: The paper strongly emphasizes that capturing sub-daily oceanic phenomena is essential for accurate forecasting, especially during extreme events. Daily-averaged models miss critical dynamics.
- Data-Driven Air-Sea Coupling: Integrating atmospheric forcings directly into a data-driven ocean model (without numerical weather models) is a significant step towards a fully AI-driven coupled prediction system.
- Multi-Timescale Dynamics: The MoT module's ability to adaptively learn and weight temporal dependencies for different variables is key to achieving robust forecasts across various timescales.
- Computational Efficiency: Data-driven models, like TianHai, offer a substantial advantage in computational cost compared to traditional numerical models, enabling faster and more scalable operational deployment.
- Vertical Resolution Matters: The model's ability to resolve ocean dynamics down to 1500m depth with 26 vertical layers is critical for capturing subsurface processes.
Key Results
- State-of-the-Art Accuracy: TianHai consistently outperforms numerical GOFS (HYCOM, OceanMAPS, FOAM) and the deep learning GOFS WenHai across multiple variables (SST, temperature/salinity profiles, currents, SLA) for lead times up to 10 days.
- Significant RMSE Reduction: TianHai reduced RMSE compared to WenHai by:
- 14.4% for SST
- 6.7% for SLA
- 8.1% for temperature profile
- 3.4% for salinity profiles
- 19.3% for 15-m zonal velocity
- 19.2% for 15-m meridional velocity
- Improved Extreme Event Forecasting: During Typhoon Aere, TianHai's 6-hour forecasts captured rapid SST changes and cooling effects more accurately than WenHai's daily forecasts.
- Enhanced Computational Efficiency: TianHai requires substantially fewer computational resources than traditional numerical models.
- Robustness Across Lead Times: The model demonstrates stable performance and slower error growth over longer forecast horizons.
Limitations and Future Work
- Reliance on Reanalysis Data: TianHai's training relies on HYCOM reanalysis data, which inherently carries biases from the underlying numerical models. Future work will focus on integrating in-situ observations to build a more comprehensive and bias-reduced training dataset.
- Longer-Term Forecasts: The current evaluation focuses on 1-10 day lead times. The model's performance on seasonal to interannual scales remains unexplored and is a key area for future research.
- Abyssal Depths: Extending vertical coverage beyond 1500m into abyssal depths would improve the representation of the full ocean system.
- Physical Constraints: Incorporating physical constraints (e.g., conservation laws) into the model architecture could further enhance prediction stability and physical consistency, particularly for fast-changing variables like currents.
- Current Velocity Prediction: While improved, current velocity predictions showed relatively muted gains compared to thermohaline variables. Future work aims to enhance this through physics-informed hybrid modeling.
Summary updated