Artificial intelligence achieves easy-to-adapt nonlinear global temperature reconstructions using minimal local data.

Area of Science:

• Climate Science
• Machine Learning
• Data Reconstruction

Background:

• Understanding climate variability is crucial for predicting future climate extremes.
• Existing climate reconstruction methods face limitations such as high costs, linearity assumptions, and sparse data.
• Climate field reconstructions and reanalyses are key tools for studying climate variability.

Purpose of the Study:

• To present a novel machine learning-based non-linear climate variability reconstruction method.
• To demonstrate the method's capability in reconstructing global temperature anomalies over 400 years.
• To assess the method's performance against established techniques.

Main Methods:

• Utilized a Recurrent Neural Network (RNN) for non-linear climate variability reconstruction.
• Trained the RNN on existing model outputs and reanalysis data.
• Applied the method to reconstruct global, monthly temperature anomalies using sparse, pseudo-station data.

Main Results:

• Successfully reconstructed over 400 years of global, monthly temperature anomalies.
• Demonstrated realistic temperature pattern and magnitude reproduction.
• Achieved reconstructions with computational costs of approximately 1 hour on a standard laptop.
• Showcased comparable performance in mean statistics to established methods.
• Highlighted suitability for reconstructing specific climate events.

Conclusions:

• The developed machine learning method offers a cost-effective and efficient approach for climate variability reconstruction.
• The Recurrent Neural Network model accurately captures temperature patterns and magnitudes, even with sparse data.
• This adaptable method can be applied to various regions, time periods, and climate variables for enhanced climate research.