Exploring dominant processes for multi-month predictability of western Pacific precipitation using deep learning

Abstract

Over the past half-century, there has been an increasing trend in the magnitude and duration of the Madden-Julian Oscillation (MJO) attributable to the significant warming trend in the Western Pacific (WP). The MJO, bridging weather and climate, influences global and regional climate through atmospheric teleconnections, and climate models can predict it for up to 4–5 weeks. In this study, we use deep learning (DL) methods to investigate the predictability of the MJO-related western Pacific precipitation on a multi-month time scale (5–9 weeks). We examine numerous potential predictors across the tropics, selected based on major MJO theories and mechanisms, to identify key factors for long-term MJO prediction. Our results show that DL-based useful potential predictability of the WP precipitation can be extended up to 6–7 weeks, with a correlation coefficient skill ranging from 0.60 to 0.65. Observational and heat map analysis suggest that cooling anomalies in the central Pacific play a crucial role in enhancing westerly anomalies over the Indian Ocean and warming in the WP, thereby strengthening the Walker circulation in the equatorial Pacific. In addition, the predictability of WP precipitation is higher in La Nina years than in El Nino or normal years, suggesting that mean cooling in the central Pacific may contribute to increased predictability of the MJO-related WP precipitation on the multi-month time scale. Additional model experiments using observed sea surface temperature (SST) anomalies over the central Pacific confirmed that these anomalies contribute to enhanced MJO-related convective anomalies over the WP. The study highlights that DL is a valuable tool not only for improving MJO-related WP prediction but also for efficiently exploring potential mechanisms linked to long-term predictability.