Exploring dominant processes for multi-month MJO prediction using deep learning

Abstract

Abstract Over a half-century, western Pacific (WP) warming trends are robust, increasing the magnitude and duration of Madden-Jullian Oscillation (MJO). The MJO affects global and regional climate through atmospheric teleconnection but the predictability of MJO in WP is limited up to 3-4weeks. Here, we utilized deep learning (DL) methods to investigate multi-month time scale (5–9 weeks) predictability. We tested many possible predictors over tropics based on major MJO theories or mechanisms to find a potential key factor for multi-month time scale MJO prediction. We showed that the potential predictability of MJO-related precipitation using DL extends to 6–7 weeks with a correlation of 0.60–0.65. The observational and heatmap analysis indicates that cooling anomalies in the central Pacific may contribute to increasing multi-month predictability by enhancing westerly anomalies over the Indian Ocean and warming in the WP with strong Walker circulation in the equatorial Pacific. Additional model experiments with observed sea surface temperature (SST) anomalies over the central Pacific (CP) confirm the contribution of CP SST to improved MJO-related convective anomalies over WP. These results show that DL is a useful tool for not only the improvement of MJO prediction but also for exploring possible mechanisms related to long-term predictability efficiently.