Machine learning helps to strongly reduce future warming uncertainty.

Area of Science:

• Climate Science
• Machine Learning Applications
• Earth System Science

Background:

• Global mean surface temperature trends are used for climate projections, but spatial warming patterns are underutilized.
• Existing methods for reducing climate projection uncertainty often overlook detailed spatial warming information.

Purpose of the Study:

• To apply machine learning to identify emergent constraints from spatial warming trends.
• To improve the accuracy of future global warming projections by incorporating historical spatial warming data.

Main Methods:

• Utilized machine learning to analyze 1971-2020 warming trends across individual grid cells in climate model simulations.
• Developed spatially resolved emergent constraint relationships between historical warming patterns and future global mean warming.

Main Results:

• Identified key tropical and polar regions where historical warming effectively constrains future global warming.
• Incorporating spatial warming patterns reduced projection error variance by approximately 70%, compared to 48% using global mean trends alone.
• Refined projections indicate a higher likelihood of exceeding Paris Agreement temperature thresholds, with an 80% chance of surpassing 2°C by mid-century under SSP3-7.0.

Conclusions:

• Machine learning can uncover physically interpretable emergent constraints from spatial climate data.
• The use of spatial warming patterns offers a significant improvement in reducing uncertainty for future climate projections.
• Enhanced climate projections suggest a greater urgency in addressing climate change to meet international targets.