Accurately predicting nitrogen (N) pollution in rivers is crucial for reducing farm-related water pollution in the Upper Mississippi River Basin, a major cause of low-oxygen 'dead zones' in the Gulf. However, current models struggle to predict daily N levels in HUC12 watersheds due to limited monitoring data and the complexity of water flow and weather impacts. To solve this, we developed a new AI model that combines graph machine learning with process-based model. Our model uses data like weather, soil, land use, and farm practices while respecting natural rules like water and mass balance. It accurately predicts daily N pollution in high spatial resolution, outperforming conventional methods. The model also pinpoints high-risk areas and times for N pollution, showing that extreme precipitations cause most of the annual N runoff. By blending AI with scientific knowledge, our approach provides a clearer, more practical tool for managing water quality. This work shows how hybrid models can help protect rivers despite data gaps and climate change.