ENSO‐Related Precursor Pathways of Interannual Thermal Anomalies Identified Using a Transformer‐Based Deep Learning Model in the Tropical Pacific

• Published in: Geophysical research letters Vol. 51; no. 12
• Main Authors: Zhou, Lu, Zhang, Rong‐Hua
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 28.06.2024; Wiley
• Subjects: Deep learning; El Nino; El Nino phenomena; El Nino-Southern Oscillation event; ENSO predictions; Equatorial regions; Evolution; explainable artificial intelligence (XAI); Initial conditions; multivariate three‐dimensional (3D) predictions; Neural networks; Neural stem cells; Ocean temperature; Oceans; Precursors; Prediction models; Southern Oscillation; Temperature anomalies; thermal precursors; Transformers; transformer‐based model
• DOI: 10.1029/2023GL107347
• ISSN: 0094-8276

Recent studies have demonstrated great values of deep‐learning (DL) methods for improving El Niño‐Southern Oscillation (ENSO) predictions. However, the black‐box nature of DL makes it challenging to physically interpret mechanisms responsible for successful ENSO predictions. Here, we demonstrate an interpretable method by performing perturbation experiments to predictors and quantifying input‐output relationships in predictions by using a transformer‐based model; ENSO‐related thermal precursors serving as initial conditions during multi‐month time intervals (TIs) are identified in the equatorial‐northern Pacific, acting to precondition input predictors to provide for long‐lead ENSO predictability. Results reveal the existence of upper‐ocean temperature anomaly pathways and consistent phase propagations of thermal precursors around the tropical Pacific. It is illustrated that three‐dimensional thermal fields and their basinwide evolution during long TIs act to enhance long‐lead prediction skills of ENSO. These physically explainable results indicate that neural networks can adequately represent predictable precursors in the input predictors for successful ENSO predictions. Plain Language Summary Deep learning (DL) methods have emerged as a powerful tool for improving El Niño‐Southern Oscillation (ENSO) predictions. But DL‐based modeling looks like “black boxes” without effectively telling why good predictions can be made. In this study, we conduct interpretable analyses to uncover the key physical processes responsible for successful ENSO predictions using a DL‐based prediction model. Results identify ENSO‐related thermal precursors in the equatorial‐northern Pacific region, which precondition ENSO evolution months ahead of time. Specifically, interannual thermal precursors are illustrated to have consistent and coherent phase propagations in the tropical Pacific basin: eastward along the equator, westward across the off‐equatorial tropical North Pacific, and apparent meridional phase connections both in the western and eastern boundaries. From the prediction perspective, the demonstrated existence of upper‐ocean temperature anomaly pathways acts to enhance long‐lead ENSO predictability in the purely data‐driven DL framework. These physically explainable results indicate that the neural networks, despite their absence of explicit physical constraints, are capable of representing predictable precursors whose information is included in the input predictors, being able to make convincing and successful ENSO predictions. Key Points A deep learning (DL) model is used to conduct El Niño‐Southern Oscillation (ENSO) predictability studies for physical interpretability DL model experiments are made to identify ENSO‐related thermal precursors along a counterclockwise pathway encircling the tropical Pacific The existence of upper‐ocean thermal anomaly pathways is demonstrated to enhance long‐lead ENSO predictability