Applying Machine Learning to Improve Simulations of a Chaotic Dynamical System Using Empirical Error Correction.

Area of Science:

• Earth System Science
• Climate Modeling
• Machine Learning Applications

Background:

• Dynamical weather and climate models are crucial for Earth system studies and future climate change projections.
• Current models exhibit discrepancies when compared to observational data.
• Machine learning (ML) has shown success in prediction tasks, prompting its consideration as a replacement for traditional models.

Purpose of the Study:

• To test a framework combining ML with physics-based models to correct model errors dynamically.
• To improve the quality of climate simulations while preserving physical principles.
• To assess the efficacy of this hybrid approach in enhancing prediction skill and climate statistics.

Main Methods:

• A novel framework was developed integrating ML with existing physics-based dynamical models.
• The ML component was trained to correct time-step errors of the physical models.
• The approach was validated using simulations of the chaotic Lorenz '96 system.

Main Results:

• The hybrid ML-physical model framework demonstrated stability in simulations.
• The approach led to improved skill in initialized predictions.
• Enhanced long-term climate statistics were observed, though improvements were smaller than for single time-step corrections.

Conclusions:

• Combining ML with physics-based models offers a promising avenue for improving climate simulations.
• The hybrid approach maintains physical understanding and reduces computational demands.
• Further research is needed to develop methods targeting improvements on longer time scales for climate projections.