Reduced cloud cover errors in a hybrid AI-climate model through equation discovery and automatic tuning

Area of Science:

• Climate Science
• Atmospheric Physics
• Machine Learning Applications

Background:

• Cloud parameterizations are a major source of uncertainty in climate projections.
• Existing data-driven machine learning approaches for Earth system models (ESMs) often lack interpretability and physical consistency.

Purpose of the Study:

• To develop and implement a physically consistent, interpretable machine-learning parameterization for cloud cover in a global atmospheric model.
• To improve the accuracy and reduce biases in climate projections by enhancing ESMs.

Main Methods:

• Incorporated a physically consistent cloud cover parameterization derived from storm-resolving simulations using symbolic regression into the ICON global atmospheric model.
• Applied the Nelder-Mead optimizer for automatic recalibration of the hybrid model against Earth observations in nested stages.

Main Results:

• The hybrid model significantly reduced biases in cloud cover, with a 75% reduction over the Southern Ocean and 44% in subtropical stratocumulus regions.
• The improved model demonstrated robustness under +4K surface warming conditions.

Conclusions:

• Interpretable machine-learned parameterizations, combined with practical tuning methods, can effectively enhance the fidelity of Earth system models.
• This approach offers a transparent and efficient way to strengthen climate projections.