Physics-informed machine learning: case studies for weather and climate modelling.

Area of Science:

• Earth Science
• Atmospheric Science
• Climate Science
• Artificial Intelligence

Background:

• Machine learning (ML) offers powerful tools for analyzing complex patterns in weather and climate data.
• Standard ML models often lack physical consistency and generalize poorly to unseen conditions.
• Integrating domain knowledge is crucial for robust ML applications in geosciences.

Purpose of the Study:

• To survey systematic approaches for incorporating physics and domain knowledge into ML models.
• To demonstrate the successful application of these physics-informed ML methods in weather and climate modeling through case studies.
• To identify challenges and future directions for developing reliable physics-informed ML for climate science.

Main Methods:

• Systematic review of approaches to integrate physical laws and domain expertise into ML algorithms.
• Analysis of 10 diverse case studies showcasing physics-informed ML in weather and climate emulation, downscaling, and forecasting.
• Synthesis of findings to categorize successful integration strategies.

Main Results:

• Physics-informed ML models exhibit improved physical consistency compared to standard ML.
• Integration of domain knowledge leads to reduced training times and enhanced data efficiency.
• These methods demonstrate superior generalization capabilities for weather and climate prediction tasks.
• Case studies highlight successful applications in emulation, downscaling, and forecasting.

Conclusions:

• Incorporating physics into ML is essential for creating accurate, reliable, and generalizable weather and climate models.
• Significant advancements have been made, but challenges remain in computational resources, diagnostics, and scientific integration.
• Future research should focus on overcoming these challenges to develop truly robust physics-informed ML for climate applications.