Climate-invariant machine learning.

Area of Science:

• Earth System Science
• Climate Modeling
• Machine Learning Applications

Background:

• Climate change projections rely on physical models that struggle with sub-grid scale processes, a key source of uncertainty.
• Machine learning (ML) offers potential for improving these process representations but often fails to generalize to unseen climate regimes.

Purpose of the Study:

• To develop a novel framework, termed "climate-invariant" machine learning (ML), that integrates physical knowledge into ML algorithms.
• To enhance the accuracy and generalizability of ML models for climate process representations across various climate conditions.

Main Methods:

• Proposed a "climate-invariant" ML framework that incorporates physical process knowledge into ML algorithms.
• Tested the framework's performance across three distinct atmospheric models.
• Evaluated the ML models' offline accuracy across a wide range of climate conditions and configurations.

Main Results:

• The climate-invariant ML framework demonstrated high offline accuracy across diverse climate conditions and model configurations.
• Explicitly incorporating physical knowledge improved the consistency and data efficiency of the ML models.
• The approach showed enhanced generalizability for data-driven models of Earth system processes.

Conclusions:

• Integrating physical knowledge into data-driven models is crucial for improving climate projections.
• The proposed climate-invariant ML framework offers a promising approach to overcome limitations in current climate modeling.
• This method enhances the reliability and applicability of ML in understanding and projecting Earth system changes.