MJO Initiation Predictability in an AI-CNOP Framework: Unraveling Precursor Signals of Primary MJO Events

The Madden–Julian Oscillation (MJO) is a major source of subseasonal-to-seasonal predictability. Identifying reliable precursor signals is crucial for improving MJO prediction and guiding model development. Given the computational burden of numerical models, we developed a data-driven model trained on the real-time multivariate MJO (RMM) index. Using the preceding 42‑day RMM index to predict the next 42 days, the model achieves a useful skill of 20 days and captures more MJO initiation events than dynamical models. This enables us to investigate optimal precursors for primary MJO initiation using the time‑extended conditional nonlinear optimal perturbation (CNOP) method. Results show that CNOPs can trigger strong MJO events from non‑MJO reference states, thus serving as precursor signals. Furthermore, the full 42‑day CNOP allows quantification of contributions from different time‑step components. Sensitivity experiments reveal an optimal triggering time that depends on the reference state. Notably, the CNOP component at the optimal timing interacts nonlinearly with residual components, from which a westward‑propagating disturbance (WPD) is extracted in 90% of CNOP‑based precursors. The WPD, identified mathematically, should play a central role in initiating primary MJO events, validating recent observational findings. These results underscore the power of the AI–CNOP framework in unraveling MJO initiation dynamics and its promise for operational subseasonal forecasting.