A self-attention-based neural network for three-dimensional multivariate modeling and its skillful ENSO predictions

Area of Science:

• Climate Science
• Oceanography
• Artificial Intelligence

Background:

• Traditional dynamical models for El Niño-Southern Oscillation (ENSO) predictions suffer from significant biases and uncertainties.
• Deep learning algorithms offer a promising avenue for improving tropical Pacific sea surface temperature (SST) modeling.

Purpose of the Study:

• To develop and evaluate a novel self-attention-based neural network, 3D-Geoformer, for enhanced ENSO predictions.
• To assess the model's capability in predicting 3D upper-ocean temperature and wind stress anomalies.

Main Methods:

• A data-driven approach utilizing a Transformer-based neural network architecture (3D-Geoformer).
• Incorporation of time-space attention mechanisms to capture complex spatiotemporal dynamics.
• Prediction of three-dimensional upper-ocean temperature anomalies and wind stress anomalies.

Main Results:

• Achieved high correlation skills for Niño 3.4 SST anomaly predictions 18 months in advance.
• Demonstrated the model's ability to capture the evolution of upper-ocean temperature and coupled ocean-atmosphere dynamics.
• Sensitivity experiments confirmed the depiction of Bjerknes feedback mechanisms during ENSO cycles.

Conclusions:

• The 3D-Geoformer model shows significant potential for improving ENSO prediction accuracy.
• Self-attention mechanisms are effective for multidimensional spatiotemporal modeling in climate science.
• This data-driven model represents a substantial advancement over traditional dynamical ENSO prediction methods.