Skillful and interpretable ENSO prediction via a glass-box dynamical-deep learning model

The El Niño–Southern Oscillation (ENSO) is a dominant mode of interannual climate variability with profound global socioeconomic and ecological impacts 1–3 . Skillful long-lead prediction of ENSO is critical but remains hindered by nonlinear ocean–atmosphere coupling and the formidable spring predictability barrier (SPB) 4–6 . While recent deep learning models have shown promise in extending ENSO forecasts beyond a lead time of one year, they largely operate as opaque “black boxes” that mask underlying physical mechanisms. Here we leverage dynamical system deep learning (DSDL) 7,8 to construct a fully transparent, multivariate “glass-box” ENSO prediction model. Our model achieves skillful predictions up to 19 months ahead, robustly circumventing the SPB and significantly outperforming state-of-the-art models. Notably, it successfully hindcasts the onset, intensity, and decay of the 2015–2016 super El Niño event more than a year in advance, even when initialized during the boreal spring. Physical interpretation reveals that the “glass-box” DSDL model functions as a multi-basin synergistic prediction framework, which further predicts a potential super El Niño development in 2026–2027. By rendering all model terms explicit, this approach transforms the “black box” into a physically accountable dynamical system, representing a fundamental advance in interpretable climate forecasting.