Predicting Atlantic and Benguela Niño events with deep learning

Extreme Atlantic and Benguela Niño events continue to significantly impact the tropical Atlantic region, with far-reaching consequences for African climate and ecosystems. Despite attempts to forecast these events using traditional seasonal forecasting systems, success remains low, reinforcing the growing idea that these events are unpredictable. To overcome the limitations of dynamical prediction systems, we introduce a deep learning-based statistical prediction model for Atlantic and Benguela Niño events. Our convolutional neural network (CNN) model, trained on 90 years of reanalysis data incorporating surface and 100m-averaged temperature variables, demonstrates the capability to forecast the Atlantic and Benguela Niño indices with lead times of up to 3-4 months. Notably, the CNN model excels in forecasting peak-season events with remarkable accuracy extending up to 5 months ahead. Gradient sensitivity analysis reveals the ability of the CNN model to exploit known physical precursors, particularly the connection to equatorial dynamics and the South Atlantic Anticyclone, for accurate predictions of Benguela Niño events. This study challenges the perception of the Tropical Atlantic as inherently unpredictable, underscoring the potential of deep learning to enhance our understanding and forecasting of critical climate events.