A spatiotemporal 3D convolutional neural network model for ENSO predictions: A test case for the 2020/21 La Nina conditions

Many coupled models are unable to accurately depict the multi-year La Nina conditions in the tropical Pacific during 2020-22, which poses a new challenge for real-time El Nino-Southern Oscillation (ENSO) predictions. Yet, the corresponding processes responsible for the multi-year coolings are still not understood well. In this paper, reanalysis products are analyzed to examine the ocean-atmosphere interactions in the tropical Pacific that have led to the evolution of sea surface temperature (SST) in the central-eastern equatorial Pacific, including the strong anomalous southeasterly winds over the southeastern tropical Pacific and the related subsurface thermal anomalies. Meanwhile, a divided temporal and spatial (TS) 3D convolution neural network (CNN) model, named TS-3DCNN, was developed to make predictions of the 2020/21 La Nina conditions; results from this novel data -driven model are compared with those from a physics-based intermediate coupled model (ICM). The prediction results made using the TS-3DCNN model for the 2020-22 La Nina indicate that this deep learning-based model can capture the two-year La Nina event to some extent, and is comparable to the IOCAS ICM; the latter dynamical model yields a successful real-time prediction of the Nino3.4 SST anomaly in late 2021 when it is initiated from early 2021. For physical interpretability, sensitivity experiments were designed and carried out to confirm the dominant roles played by the anomalous southeasterly wind and subsurface temperature fields in sustaining the second-year cooling in late 2021. As a potential approach to improving predictions for diversities of ENSO events, additional studies on effectively combining neural networks with dynamical processes and mechanisms are expected to significantly enhance the ENSO prediction capability.