Deep learning to represent subgrid processes in climate models.

Area of Science:

• Climate modeling
• Atmospheric science
• Artificial intelligence

Background:

• Nonlinear subgrid processes, particularly clouds, introduce significant uncertainty in climate models.
• Cloud-resolving models offer better process representation but are computationally limited to short simulations.

Purpose of the Study:

• To demonstrate the feasibility of using deep learning to emulate cloud-resolving model capabilities at reduced computational cost.
• To integrate a deep learning-based parameterization into a global general circulation model.

Main Methods:

• A deep neural network was trained on data from a multiscale model with explicit convection.
• The trained neural network replaced traditional subgrid parameterizations in a global climate model.
• Multiyear simulations were conducted to evaluate model performance.

Main Results:

• The deep learning parameterization reproduced the mean climate and variability (precipitation extremes, equatorial waves) of the cloud-resolving simulation.
• The neural network achieved approximate energy conservation without explicit instruction.
• The parameterization generalized to new surface forcing but had limitations with out-of-distribution temperatures.

Conclusions:

• Deep learning is a viable approach for climate model parameterization.
• Data-driven methods are expected to reduce climate prediction uncertainty.
• This work paves the way for more efficient and accurate Earth system models.