Coupled Ocean-Atmosphere Dynamics in a Machine Learning Earth System Model (2406.08632v1)

Abstract: Seasonal climate forecasts are socioeconomically important for managing the impacts of extreme weather events and for planning in sectors like agriculture and energy. Climate predictability on seasonal timescales is tied to boundary effects of the ocean on the atmosphere and coupled interactions in the ocean-atmosphere system. We present the Ocean-linked-atmosphere (Ola) model, a high-resolution (0.25{\deg}) Artificial Intelligence/ Machine Learning (AI/ML) coupled earth-system model which separately models the ocean and atmosphere dynamics using an autoregressive Spherical Fourier Neural Operator architecture, with a view towards enabling fast, accurate, large ensemble forecasts on the seasonal timescale. We find that Ola exhibits learned characteristics of ocean-atmosphere coupled dynamics including tropical oceanic waves with appropriate phase speeds, and an internally generated El Ni~no/Southern Oscillation (ENSO) having realistic amplitude, geographic structure, and vertical structure within the ocean mixed layer. We present initial evidence of skill in forecasting the ENSO which compares favorably to the SPEAR model of the Geophysical Fluid Dynamics Laboratory.

• The paper introduces the Ola model, an ML-driven coupled system that delivers realistic ENSO forecasts and reduces tropical biases.
• The model accurately simulates tropical SST variability while generating equatorial Kelvin and Rossby waves with realistic phase speeds.
• The paper demonstrates the potential of ML approaches to enhance seasonal climate predictions and reduce errors in operational models.

Coupled Ocean-Atmosphere Dynamics in a Machine Learning Earth System Model

The research paper presents the Ocean-linked-atmosphere (Ola) model, a high-resolution AI/ML coupled Earth-system model designed for seasonal climate forecasting. This model operates at a $0.25^\circ$ resolution, leveraging an autoregressive Spherical Fourier Neural Operator (SFNO) architecture. Ola separately models ocean and atmosphere dynamics, offering a promising approach to fast, accurate, and large-scale ensemble forecasts on seasonal timescales.

Key Findings

1. ENSO Forecasting: Ola exhibits skill in forecasting the El Niño/Southern Oscillation (ENSO) by generating realistic ocean-atmosphere dynamics, including tropical oceanic waves with appropriate phase speeds. The model's simulations produce ENSO events with realistic amplitude, geographic structure, and vertical structure within the ocean mixed layer, comparing favorably to the SPEAR model from the Geophysical Fluid Dynamics Laboratory.
2. Tropical SST Variability: Over a five-year validation period (2017-2022), Ola generated realistic central tropical Pacific Sea Surface Temperature (SST) variability, accurately producing both El Niño and La Niña states. Although there was an exception during the persistent La Niña events in 2020/2021 and 2021/2022, this anomaly was consistent with the difficulty observed in other operational models.
3. Systematic Bias Reduction: Unlike the GFDL-SPEAR model, Ola presents significantly lesser time-mean cold bias in the Niño 3.4 region, indicating the potential of ML models in mitigating tropical drifts often seen in traditional process-based models. This reduction in bias could provide better ENSO forecasting and reduce associated errors in large-scale atmospheric circulation and hydrological cycles.
4. Oceanic Wave Dynamics: Ola successfully generates oceanic equatorial Kelvin and Rossby waves, critical for ENSO dynamics. The forecasts qualitatively exhibit eastward and westward propagating SSH anomalies with realistic phase speeds, confirming the model's capability to simulate essential oceanic wave dynamics.
5. ENSO Ocean Temperature Structure: The model accurately reproduces the three-dimensional thermal structure of ocean temperature anomalies during ENSO events. Simulated composites of El Niño and La Niña events closely match observed data, showcasing Ola's proficiency in modeling both surface and subsurface temperature anomalies.
6. Transient Dynamics: Case studies on the onset of El Niño and La Niña events demonstrated Ola’s ability to simulate the plausible transient evolution of these phenomena. The model captures complex interactions between tropical and subtropical controls, providing insights into the varying pathways that lead to ENSO development.

Implications

Practical and Theoretical Impact

The Ola model's ability to produce accurate seasonal forecasts in a fraction of the time required by traditional models presents significant practical advantages. This efficiency, combined with the model's reduced bias and realistic wave dynamics, opens avenues for improved operational climate prediction and risk management in sectors such as agriculture and energy.

From a theoretical perspective, this paper emphasizes the potential of ML models to learn complex coupled atmosphere-ocean dynamics without the biases inherent in process-based models. The model's success in simulating realistic ENSO characteristics and oceanic waves suggests that ML approaches could eventually circumvent the need for complex subgrid parameterizations, traditionally necessary in physical climate models.

Future Directions

The paper identifies several areas worth exploring to enhance coupled AI/ML models:

1. Extended Training Data: Developing and utilizing extensive datasets spanning thousands of years of climate model simulations would be beneficial to better capture internal variability and forced dynamics, crucial for long-term climate predictions.
2. Incorporating Enhanced Ocean Dynamics: Future models could feature more detailed subsurface ocean state vectors, potentially improving the representation of slower ocean circulation components, such as subtropical gyres and the Atlantic meridional overturning circulation.
3. Stability and Drift Control: Addressing the inherent long-term drift issues seen at higher latitudes in Ola is crucial. Understanding and mitigating these drifts through systematic empirical testing and model configuration optimization could lead to more stable long-term forecasts.
4. Operational Integration: Continuous evaluation and incremental development are necessary before Ola can be applied in operational settings. Formal comparisons with existing operational models would help benchmark its relative performance and identify areas for improvement.

Conclusion

The Ola model represents a significant step towards integrating ML into coupled atmosphere-ocean forecasting. With its demonstrated ability to simulate realistic ENSO dynamics and reduced tropical biases, Ola showcases the potential of AI/ML models in advancing climate science. Future work focused on extending training datasets, controlling forecast drift, and improving model resolution can further enhance the accuracy and reliability of these models for operational use.

By making the code, model weights, and data publicly available, this research also lays the groundwork for collaborative advancements in the field, encouraging a transition from medium-range weather forecasting to robust seasonal and potentially longer-term climate predictions utilizing ML methodologies.