Deep learning trained on high-fidelity climate model simulations extends the skillful prediction of Northern Tropical Atlantic SST anomalies

Abstract

Boreal spring sea surface temperature (SST) anomalies in the northern tropical Atlantic (NTA) are key drivers of global climate extremes, yet their prediction remains challenging for state-of-the-art dynamical models. Deep learning offers promise for seasonal SST prediction but is limited by data scarcity and weak physical interpretability. Here, we develop a convolutional neural network (CNN) framework for NTA prediction and demonstrate that physics-informed training data selection improves performance. Specifically, selecting CMIP6 models that realistically capture the ENSO–NTA teleconnection, rather than using all the CMIP6 models, significantly enhances predictive skill. The CNN trained on these selected models extends effective predictions to 8 lead months, compared to 7 months in a multi-model ensemble of dynamic models, with notable improvements in boreal winter and spring. It also slightly outperforms dynamical models in predicting extreme spring NTA events. Further analysis shows that the model captures physically consistent precursors, with predictability governed by the interplay between ENSO forcing and local ocean–atmosphere variability. These results highlight the value of physics-informed data selection in deep learning and its potential to improve prediction of NTA SST anomalies.