Presented on January 11, 2023, by
Professor Daehyun Kim
Department of Atmospheric Sciences, Univ. of Washington

Abstract:

The Madden-Julian Oscillation (MJO) is the dominant mode of tropical intraseasonal variability that interacts with many other Earth system phenomena, including high-impact weather events in the midlatitude. The prediction skill of the MJO in many operational models is lower than potential predictability, partly because our understanding of the MJO’s predictability is limited.

In this study, we investigate the source of the MJO’s predictability by combining a machine learning (ML) technique with a 1200-year-long simulation made with Community Earth System Model version 2 (CESM2). A convolutional neural networks (CNN)-based MJO prediction model is first trained using the CESM2 simulation data and then fine-tuned using observational data via the transfer learning, with five 2-D fields of atmospheric variables as input and the real-time multivariate MJO indices as the output. The source of MJO predictability in the CNN model is examined via explainable artificial intelligence (XAI) methods that quantify the relative importance of the input variable.
Our CNN model outperforms previous statistical models and many operational forecasts with the prediction skill of about 25 days. Applying the XAI methods to the CNN model highlights precipitable water anomalies over the Indo-Pacific warm pool as key precursors of the subsequent MJO development for 1-3 weeks forecast lead times. Surface temperature anomalies are also found to play an important role, especially for longer (> 3 weeks) forecast lead times. Our results suggest a realistic representation of moisture dynamics is crucial for accurate MJO prediction.