Projection of ENSO using observation-informed deep learning.

Area of Science:

• Climate Science
• Machine Learning Applications
• Oceanography

Background:

• The El Niño-Southern Oscillation (ENSO) significantly influences global climate patterns.
• Existing climate models exhibit considerable uncertainty and inter-model spread in projecting future ENSO sea surface temperature (SST) variability.
• Discrepancies between climate models and observations in ENSO physics hinder accurate future projections.

Purpose of the Study:

• To reduce uncertainty in 21st-century ENSO SST variability projections.
• To develop a method for constraining climate model projections using observational data.
• To investigate the role of deep learning in improving climate model accuracy.

Main Methods:

• Utilized deep learning, specifically artificial neural networks (ANNs), trained on climate model simulations and observational data.
• Employed interpretability analyses to understand how ANNs capture ENSO physics.
• Applied a model-as-truth approach to validate ANN-generated projections.
• Conditioned future ENSO SST variability projections based on ANN-inferred ENSO responses to tropical Pacific warming.

Main Results:

• Deep learning reduced ENSO SST variability projection uncertainty by 54% under a high-emission scenario.
• ANNs successfully replicated observed ENSO responses, identifying key warming patterns in the tropical Pacific.
• The study narrowed the projected ENSO SST variability range from 0.59°C to 0.27°C.
• Interpretability confirmed that replicating observed ENSO physics was crucial for reducing uncertainty.

Conclusions:

• Integrating machine learning with observational data offers a promising approach to reduce uncertainty in climate projections.
• The developed deep learning method provides a robust constraint for refining future ENSO predictions.
• Accurate representation of ENSO physics is critical for reliable climate change projections.