Physics‐Constrained Network for Enhanced Extended‐Range Precipitation Forecasting in East Asia

The complex summer weather in East Asia poses significant challenges of precipitation forecasting. Despite the promising capabilities of deep learning models in conventional forecasting tasks, their black‐box nature limits physical interpretability. In this study, we developed PS2F‐Net, a physics‐constrained hierarchical spatio‐temporal feature fusion network. The model integrates multi‐scale physical features through a multi‐head attention mixing block and embeds the vertically integrated water vapor flux term as a physical constraint in the loss function. This physics‐informed design enhances precipitation classification and maintains forecast stability over extended lead times. Compared with the ECMWF IFS, PS2F‐Net substantially improves precipitation forecast accuracy and can predict heavy rainfall events up to 2 weeks in advance, including both typical monsoonal and typhoon‐induced extremes. Feature attribution analyses highlight the importance of nonlinear relationships in improving forecast skill. These results provide new insights and methodological support for physics‐guided deep learning in East Asian precipitation prediction.