Skillful Global Ocean Emulation and the Role of Correlation-Aware Loss
Summary: arXiv:2604.18727v1 Announce Type: cross
Abstract: Machine learning emulators have shown extraordinary skill in forecasting atmospheric states, and their application to global ocean dynamics offers similar promise. Here, we adapt the GraphCast architecture into a dedicated ocean-only emulator, driven by prescribed atmospheric conditions, for medium-range predictions. The emulator is trained on NOAA’s UFS-Replay dataset. Using a 24 hour time step, single initial condition, and without using autoregressive training, we produce an emulator that provides skillful forecasts for 10-15 day lead times. We further demonstrate the use of Mahalanobis distance as loss that improves the forecast skill compared to the Mean Squared Error loss by explicitly accounting for the correlations between tendencies of the target variables. Using spatial correlation analysis of the forecasted fields, we also show that the proposed correlation-aware loss acts as a statistical-dynamical regularizer for the slow, correlated dynamics of the global oceans, offering a better background forecast for downstream tasks like data assimilation.
Introduction
The advancement of machine learning techniques has significantly impacted various fields, including meteorology and oceanography. Among these developments, machine learning emulators have emerged as powerful tools in forecasting atmospheric states. This article explores the adaptation of machine learning emulators for global ocean dynamics, focusing on the effectiveness of correlation-aware loss functions.
The GraphCast Architecture
The GraphCast architecture, initially designed for atmospheric predictions, serves as the foundation for our ocean-only emulator. This innovative approach allows us to incorporate prescribed atmospheric conditions, enhancing our ability to generate medium-range ocean forecasts. The adaptation of this architecture is crucial for addressing the unique challenges posed by ocean dynamics.
Training and Methodology
Our ocean emulator is trained using NOAA’s UFS-Replay dataset, which provides a comprehensive set of atmospheric data. The training process involves:
- Utilizing a 24-hour time step to capture short-term oceanic changes.
- Employing a single initial condition to maintain consistency across forecasts.
- Omitting autoregressive training to streamline the forecasting process.
As a result, the emulator can generate skillful forecasts for lead times ranging from 10 to 15 days, demonstrating the potential for long-term predictions in ocean dynamics.
Correlation-Aware Loss Function
A significant contribution of this study is the introduction of a correlation-aware loss function based on Mahalanobis distance. This approach offers several advantages:
- Improved forecast skill by explicitly accounting for correlations between target variables’ tendencies.
- Enhanced ability to model the slow, correlated dynamics characteristic of global oceans.
- Functioning as a statistical-dynamical regularizer, it provides a more robust background forecast for downstream tasks, such as data assimilation.
Spatial Correlation Analysis
The effectiveness of the correlation-aware loss function is further validated through spatial correlation analysis of the forecasted fields. This analysis highlights the improved performance of the emulator and its capability to capture the intricate dynamics of ocean systems. The results indicate that incorporating correlation-aware methodologies can lead to more accurate and reliable forecasts.
Conclusion
This study emphasizes the promising role of machine learning emulators in ocean dynamics, particularly through the adaptation of the GraphCast architecture and the implementation of correlation-aware loss functions. As research continues to evolve, these approaches may significantly enhance our understanding and forecasting of global ocean behavior.