Stable machine-learning parameterization of subgrid processes for climate modeling at a range of resolutions

Area of Science:

• Climate modeling
• Atmospheric science
• Machine learning applications

Background:

• Subgrid parameterizations in global climate models introduce significant uncertainty in climate projections.
• Machine learning offers a promising avenue for developing novel parameterizations, but faces challenges with stability and climate drift.

Purpose of the Study:

• To investigate the performance of machine-learned parameterizations across different grid spacings.
• To address issues of instability and climate drift in machine-learned parameterizations.

Main Methods:

• Utilized a random forest algorithm to learn a parameterization from coarse-grained output of a high-resolution idealized atmospheric model.
• Evaluated parameterization performance at various coarse-graining factors and horizontal grid spacings.

Main Results:

• The learned parameterization resulted in stable coarse-resolution simulations that accurately replicated the high-resolution climate.
• Parameterization performance was optimal at smaller horizontal grid spacings, indicating scale-dependent effectiveness.

Conclusions:

• Machine-learned parameterizations can achieve stable and accurate climate simulations.
• Understanding parameterization performance across different scales is crucial for improving climate projections.
• This approach shows potential for developing parameterizations from emerging global high-resolution simulations.