Improving S2S Precipitation Forecast over China via a Deep Learning Model with Multi-Sphere Causality-Linked Predictors

Numerical models face persistent challenges in subseasonal-to-seasonal (S2S) precipitation forecasting over China due to strong precipitation variability and complex multi-sphere coupling at S2S timescales. In recent years, artificial intelligence (AI)-based post-processing has emerged as a promising approach, owing to its capacity to learn complex nonlinear relationships and correct systematic model biases from historical data. However, most existing AI-based methods neglect the spatial structure and physical interactions among multi-sphere predictors (e.g., atmosphere, ocean, and land), limiting their ability to capture the underlying dynamics required for physical consistency. This study develops an S2S precipitation bias-correction network (S2SPre-BCNet) based on a cycle-consistent generative adversarial network (CycleGAN), which incorporates causality-selected multi-sphere predictors as conditional inputs to improve weekly accumulated precipitation forecasts from the ECMWF S2S system over China at lead times of 1–6 weeks. Compared to the ECMWF S2S, S2SPre-BCNet reduces mean RMSE (root mean square error) by 11.6% (maximum 17.2%), increases mean ACC (anomaly correlation coefficient) by 27.2% (maximum 49.2%), and raises mean HSS (Heidke skill score) by 1.23% (maximum 2.12%). Across the case studies, S2SPre-BCNet lowers the absolute mean precipitation error by 16.4%. Additionally, interpretability analyses reveal that multi-sphere predictors contribute distinctly across lead times, and the model focuses on physically meaningful regions where precipitation dynamics are most complex, highlighting the potential of causality-informed AI for operational S2S bias correction. This study underscores that AI techniques augmented by causality-based predictor selection can effectively correct biases in forecasts produced by numerical models, enabling their use in operational forecasting.