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

Accurate subseasonal-to-seasonal (S2S) precipitation forecasting over China poses significant challenges owing to its complexity and variability. Considering the inherent errors in numerical models, post-processing their outputs emerges as an effective strategy. In recent years, post-processing methods based on artificial intelligence (AI) techniques have demonstrated considerable potential for further improving the S2S precipitation forecast over China. However, most existing AI-based post-processing methods overlook the spatial structure and physical interactions of forecast variables, especially those among multiple spheres (e.g., atmosphere, ocean, land), leading to an inability to capture the underlying physical dynamics critical for maintaining physical consistency. This study proposes an S2S precipitation-bias correction network (S2SPre-BCNet) based on the cycle-consistent generative adversarial network (CycleGAN), which is an advanced AI framework that can accommodate multi-sphere key influencing predictors selected by a causal discovery method as conditional inputs, to improve weekly accumulated precipitation forecasts from the ECMWF S2S system over China at lead times of 1-6 weeks. We evaluate the model performance using three metrics: root mean square error (RMSE), anomaly correlation coefficient (ACC), and Heidke skill score (HSS), and conduct three case studies in summer, autumn, and winter. Compared to the ECMWF S2S, S2SPre-BCNet reduces mean RMSE by 11.6%, with a maximum decrease of 17.2%; increases mean ACC by 27.2%, with a maximum improvement of 49.2%; and raises mean HSS by 1.23%, with a maximum improvement of 2.12%. Across the case studies, it lowers the absolute mean precipitation error by 16.4%. Additionally, interpretability analysis indicates that incorporating variables from different spheres with causality-based predictor selection has a physically meaningful impact on forecast results at different lead times, and that the model incorporating multi-sphere causality-selected predictors is capable of effectively capturing key regional features. 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.