Achieving Explainable ENSO Prediction Using Small DataTraining

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 biasesin 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 resultssuggest that training models with physical‐informed data is an effective approach for ENSO prediction.

KEY POINTS: 1)A hybrid model is constructed by integrating a deep learning model with a dynamical model.
2)Enhanced prediction skills arise primarily from improved representation of ENSO characteristics and the leading dynamical feedbacks associated with ENSO evolution.
3)The hybrid model requires only a small volume of training data.

The El Niño–Southern Oscillation (ENSO) is the Earth's largest source of year‐to‐year climate variability and greatly impacts global climate. Scientists have been trying to predictENSO's timing and strength, but traditional dynamical models still struggle to do this accurately. Deep learning models have shown promise, but they suffer from two major limitations: they rely on biased climate model data for training, and their predictions can be difficult to  interpret physically. To address these issues, we developed a new hybrid model that integrates a deep learning module into a dynamical module, combining the strengths of both approaches. Our model makes more accurate predictions of ENSO events. The deep learning module is trained on physically informed data generated by the dynamical module, which reduces reliance on large volumes of observational data and minimizes errors from imperfect climate simulations. The improved prediction skill primarily arises from the model's enhanced ability to represent the key physical feedbacks that drive ENSO. This work demonstrates that integrating artificial intelligence with physical science leads to more accurate and more trustworthy climate predictions.