Deep learning reveals moisture as the primary predictability source of MJO

Abstract

The Madden-Julian Oscillation (MJO) is the dominant mode of tropical intraseasonal variability that interacts with many other Earth system phenomena. The prediction skill of the MJO in many operational models is lower than its potential predictability, partly due to our limited understanding of its predictability source. Here, we investigate the source of MJO predictability by combining machine learning (ML) with a 1200-year-long Community Earth System Model version 2 (CESM2) simulation. A Convolutional Neural Network (CNN) for MJO prediction is first trained using the CESM2 simulation and then fine-tuned using observations via transfer learning. The source of MJO predictability in the CNN is examined via eXplainable Artificial Intelligence (XAI) methods that quantify the relative importance of the input variables. Our CNN exhibits an enhanced prediction skill over previous ML models, achieving a skill level of about 25 days. This level of performance is slightly superior or comparable to most operational models participating in the S2S project, although a few dynamical models surpass it. The XAI methods highlight precipitable water anomalies over the Indo-Pacific warm pool as the primary precursors of the subsequent MJO development for 1-3 weeks forecast lead times. Our results suggest that realistic representation of moisture dynamics is crucial for accurate MJO prediction.