Achieving explainable ENSO prediction using small data training

Despite substantial progress over the past four decades, accurately predicting the spatiotemporal structure of the El Niño–Southern Oscillation (ENSO) remains a persistent challenge for dynamical models. While deep learning models have demonstrated improved prediction skills, their performances are constrained by biases in climate models used for training and lack dynamic interpretability. Here we construct a novel hybrid model that integrates deep learning techniques into a dynamical model, enabling information exchanging during integration. Training on physical-informed data, the model continuously adapts and improves forecasts, achieves unprecedented ENSO prediction skills, particularly in El Niño diversity and the spring predictability barrier. Moreover, as the hybrid model requires only a small volume of data by training on observations, it circumvents biases in climate models. Enhanced prediction skills arise primarily from improved representation of the leading feedbacks associated with ENSO. Our results suggest that training models with physical-informed data is an effective approach for ENSO prediction.