Extending ENSO Prediction beyond 18 Months through Stage-consistent Ocean Internal Dynamics Based on Deep Learning

Deep learning (DL) has emerged as a transformative tool in atmospheric science, demonstrating particular efficacy in El Niño-Southern Oscillation (ENSO) prediction by achieving accurate forecasts with a notable lead time. Nevertheless, the adoption of DL-based ENSO forecasting models remains constrained by insufficient physical interpretability. In this study, we leveraged the linkage between sea subsurface temperature (SUBT) propagation and ENSO development—namely, a propagation loop in equatorial/extra-equatorial Pacific subsurface temperatures where anomaly signals propagate over timescales matching ENSO interannual recurrence—to develop two physically constrained, lightweight, and generalizable DL models for ENSO prediction: (i) a Convolutional Neural Network (CNN) architecture integrated with a Convolutional Block Attention Module (CBAM), and (ii) a ConvTransformer model built upon this CNN framework. These models effectively reconstruct SUBT anomaly propagation and skillfully forecast the Niño3.4 index at 18-month lead times. Guided by the physical interpretation of the ConvTransformer model using Integrated Gradients, we proposed a novel ENSO stage partitioning scheme that is physically consistent with the propagation of SUBT anomaly, providing a robust explanation for the onset, evolution, and transition of El Niño and La Niña events within a unified dynamical framework. This DL-derived scheme not only enhances forecast reliability but also offers new insights into the subsurface-driven mechanisms underlying ENSO variability and predictability.