The Predictor of Multi-year ENSO Forecasts Revealed by Deep Learning

The El Niño/Southern Oscillation (ENSO), characterized by the quasi-periodic but irregular, anomalous warming of the sea surface temperature in the tropical Pacific during El Niño and cooling during La Niña, significantly influences the Earth’s year-to-year climate variability. Early ENSO forecasting is of great scientific and social importance. Recent decades have evidenced the significant advancements of ENSO prediction with skillful forecasts achievable up to six months using statistical, physical, and data-driven deep learning methods. However, both statistical and physical approaches exhibit low predictability across boreal spring (February to May), constituting the spring predictability barrier in ENSO prediction. Despite numerous efforts to understand the spring predictability barrier from dynamic to numerical perspectives, the skillful forecasting of ENSO is limited to less than 12 months. In contrast, deep learning methods can cross the spring predictability barrier, providing forecasts exceeding one year. However, deep learning methods act as a “black-box”. The knowledge enabling the deep learning models to cross the spring predictability barrier remains unknown.

To uncover this hidden information, we propose a hybrid approach that integrates climate dynamical analysis with deep learning models, aiming to extract the meaningful and understandable knowledge from vast datasets. Based on this approach, we show that the hidden knowledge enabling deep learning models to cross the ENSO’s spring predictability barrier is the tropical Pacific Ocean heat uptake. Our approach starts with exploring the hidden information within deep learning models, revealing new insights into ENSO dynamics and offering novel pathways to improve future climate forecasts.