HydroTrace: an interpretable attention-based AI for glacio-hydrological modeling

Abstract Hydrological modeling is essential for effective water resource management, especially in complex regions such as the Tibetan Plateau, where conventional models often struggle to capture dynamic glacio-hydrological processes. This study introduces HydroTrace, an algorithm-driven model with custom attention mechanisms that consistently outperforms existing models across hydrological benchmarks at two observational sites on the Tibetan Plateau. HydroTrace achieves high predictive accuracy, with a calibrated Nash-Sutcliffe Efficiency (NSE) of 0.98, enables direct mapping of input–output contributions, identifies key physical drivers, and uses learned attention to guide water-balance regression. The approach reveals a previously overlooked year-round glacier influence on river discharge across the central and southeastern Plateau: in the downstream southeast, glaciers, melt, and baseflow co-vary with the monsoon yet glaciers remain a meaningful contributor outside the ablation peak; at higher central elevations, hydrology remains glacier-dominated throughout the year. By bridging algorithm-driven and physics-based modeling, with attention serving as a data-informed proxy, HydroTrace offers a practical pathway toward high-performing, interpretable, process-aware hydrological AI.