Deep learning to represent sub-grid processes in climate models (1806.04731v3)

Abstract: The representation of nonlinear sub-grid processes, especially clouds, has been a major source of uncertainty in climate models for decades. Cloud-resolving models better represent many of these processes and can now be run globally but only for short-term simulations of at most a few years because of computational limitations. Here we demonstrate that deep learning can be used to capture many advantages of cloud-resolving modeling at a fraction of the computational cost. We train a deep neural network to represent all atmospheric sub-grid processes in a climate model by learning from a multi-scale model in which convection is treated explicitly. The trained neural network then replaces the traditional sub-grid parameterizations in a global general circulation model in which it freely interacts with the resolved dynamics and the surface-flux scheme. The prognostic multi-year simulations are stable and closely reproduce not only the mean climate of the cloud-resolving simulation but also key aspects of variability, including precipitation extremes and the equatorial wave spectrum. Furthermore, the neural network approximately conserves energy despite not being explicitly instructed to. Finally, we show that the neural network parameterization generalizes to new surface forcing patterns but struggles to cope with temperatures far outside its training manifold. Our results show the feasibility of using deep learning for climate model parameterization. In a broader context, we anticipate that data-driven Earth System Model development could play a key role in reducing climate prediction uncertainty in the coming decade.

Deep Learning for Sub-Grid Parameterization in Climate Models

This paper presents a novel application of deep learning to enhance the parameterization of sub-grid processes in climate models, particularly focusing on clouds. Traditional climate models process atmospheric phenomena at resolutions around 50-100 km, which fails to explicitly resolve crucial small-scale processes like cloud formation and atmospheric convection. Historically, these have been parameterized using heuristic methods, but inaccuracies persist, especially given clouds' significant impact on climate dynamics.

The authors propose an alternative approach by leveraging deep learning to create a more efficient and precise parameterization. They train a deep neural network on data derived from cloud-resolving models (CRMs), which offer a higher resolution and more accurate representation of such processes but are computationally prohibitive on a global scale for long-term simulations.

Key Findings

The neural network replaces traditional parameterizations within a global general circulation model (GCM) and demonstrates several significant outcomes:

1. Cost Efficiency: The deep learning model is computationally more efficient—a key advantage, as the neural network executes approximately 20 times faster than the CRM-based model, offering a path forward for scalable computations in climate modeling.
2. Climate Reproduction: It successfully reproduces mean climate states and crucial variability aspects of the CRM, including precipitation extremes and the equatorial wave spectrum.
3. Energy Conservation: The neural network approximately conserves energy, even though it wasn't explicitly constrained to do so during training, indicating an implicit learning of some physical principles.
4. Generalization: While the model generalizes well to different surface forcing patterns, it struggles when faced with temperature regimes significantly outside its training data.

Implications

This work suggests deep learning could substantially reduce uncertainties in climate predictions by improving the representation of complex sub-grid processes within Earth System Models. The ability of neural networks to extrapolate complex dynamics and adapt to new forcing scenarios could lead to more reliable climate projections.

Future Directions

Moving forward, several challenges and potential pathways are noted:

• Complexity and Stability: Incorporating additional complexity, like ensuring positivity of water concentrations and momentum transport, and stabilizing these enhancements is essential.
• Generalization: Developing techniques to improve network generalization to out-of-sample climates is critical. Combining neural networks with established physics-based approaches may enhance robustness and forecast reliability.
• Real-World Applications: A significant aim is to transition this approach from idealized models to more comprehensive real-world applications, considering factors like topography and aerosols. Moreover, the pursuit of using observational data to train networks is highlighted, though current limitations in data availability and quality pose challenges.

The paper embraces a paradigm shift towards a data-driven model development approach, potentially transforming how sub-grid processes are represented in climate models. Emphasizing the integration of machine learning with traditional methodologies could open new vistas for enhancing climate model fidelity and accuracy. The continuation of research in this domain promises substantial contributions to the evolving landscape of climate science and modeling.