Subseasonal-to-seasonal (S2S) prediction on the 2–8-week timescale plays a vital role in food security, water management, and disaster risk reduction. By learning nonlinear relationships directly from observational and reanalysis data, artificial intelligence (AI) has emerged as an important technical pathway for advancing S2S prediction. This paper reviews recent progress along two complementary directions. Machine-learning post-processing of dynamical outputs effectively corrects forecast biases in temperature and precipitation. Data-driven models, in parallel, have evolved from target-specific systems to global multi-variable architectures, and further to multi-sphere fully coupled Earth system models, showing particularly strong skill for large-scale circulation such as the Madden–Julian Oscillation (MJO). Beneath this progress, however, forecast skill across current architectures deteriorates sharply beyond 4–5 weeks, and AI models systematically underestimate record-breaking extremes. Breaking this ceiling will require multi-source data fusion to recover tail information, probabilistic objective functions optimized with embedded physical constraints to preserve extreme signals, and interpretable architectures to support operational trustworthiness—ultimately moving toward observation-driven, end-to-end Earth system intelligent prediction with multi-sphere coupling.

Based on the persistent heavy rainfall dataset produced by China Meteorological Administration (CMA), the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center daily precipitation and Outgoing Longwave Radiation (OLR) dataset, and the ERA-Interim reanalysis product, this study analyzes relative importance of two types of low-frequency oscillations, i.e., 10—20 d and 30—60 d oscillations, for persistent heavy rainfall events during the pre-flood period over southern China. The associated mechanisms are diagnosed using the scale decomposed moisture equation and vertical velocity equation. The results reveal that the 10—20 d quasi-biweekly oscillation has a more significant impact on the intensity of persistent heavy rainfall events, while the 30—60 d intra-seasonal oscillation shows a higher correlation with the duration of heavy rainfall events. This result suggests that persistent heavy rainfall events are closely related to the occurrence and evolution of low-frequency precipitation anomalies. Based on the moisture budget diagnosis, it is found that the 10—20 d precipitation anomaly mainly comes from horizontal moisture advection induced by the interaction between wind perturbation and background moisture. For the 30—60 d precipitation anomaly, both the moisture advection and convergence processes make positive contributions to the accumulation of moisture anomalies, although advection process related to the interaction between background wind fields and 30—60 d moisture perturbation is the primary contributor. The upward motion anomaly, which results from vertical gradient of vorticity advection by the background wind–vorticity perturbation interaction, provides a favorable dynamic condition for the occurrences of both 10—20 d and 30—60 d precipitation anomalies. The above results suggest that better understanding of the scale interaction processes between low-frequency oscillations and background mean state in numerical models is the basis for accurate forecast of persistent heavy rainfall.