The Hidden Predictor of Multi-Year ENSO Predictions Revealed by Deep Learning

Abstract

Data-driven deep learning models (DLMs) can predict the El Niño-Southern Oscillation (ENSO) up to a year in advance, a capability that traditional physical and statistical models struggle to achieve due to the Spring Predictability Barrier (SPB). However, the specific knowledge that DLMs learn to cross the SPB in ENSO predictions remains unclear. In this study, we propose a non-parametric AI interpretability approach based on the extent to which useful predictable information can be include in data. By strategically reducing the training data sets to the key state required to maintain prediction skill, we uncover the critical knowledge learned by the DLMs and identify the tropical Pacific Ocean mode (TPOM) of Empirical Orthogonal Functions related to subsurface ocean temperature. The coupled ocean-atmosphere dynamics induced by TPOM are beneficial for enhancing ENSO prediction skill beyond 1 year. We integrate physical analysis with the flexibility of deep learning to reveal hidden dynamics, improve ENSO predictions, and demonstrate broad applicability in both climate science and AI interpretability.