Deep learning reveals enhanced ENSO predictability under historical anthropogenic forcing

Understanding how ENSO predictability responds to climate change is essential for improving future climate projections. In this study, we apply a deep learning model—convolutional neural network with a leave-one-out model strategy to Coupled Model Intercomparison Project Phase 6 (CMIP6) historical and preindustrial control simulations. We find that El Niño–Southern Oscillation (ENSO) predictability is statistically enhanced by 14.0 ± 1.8% enhancement under historical anthropogenic forcing. This improvement is linked to changes in key ocean-atmosphere feedbacks. Under historical forcing, the equatorial Pacific shows a shoaling of the thermocline. As a result, the surface ocean becomes more sensitive to wind forcing, and subsurface temperature becomes more responsive to thermocline variations. These processes strengthen the three-dimensional advective and thermocline feedbacks, which together enhance ENSO predictability. These findings highlight the importance of anthropogenic forcing in shaping ENSO predictability in a warming climate.