A dynamic-gated Mixture-of-Experts framework improves and interprets daily streamflow simulation

Abstract Accurately forecasting daily streamflow remains challenging as existing models struggle to balance flexibility for rapid process shifts with physical consistency. Here we present HydroMoE, a dynamic-gated Mixture-of-Experts framework pairing process-based and neural experts for each hydrological subprocess. Meteorology-responsive gates learn to allocate weight between physical and neural experts, advancing hybrid modelling toward interpretable process understanding. Across 550 CAMELS-United States basins, HydroMoE substantially improves daily streamflow prediction. In the independent test period, HydroMoE achieved a median NSE of 0.663 and a median KGE of 0.638, compared with 0.127 and 0.210 for a differentiable process-based baseline and single-module neural ablations. The learned gating weights exhibit physically interpretable patterns: runoff experts are activated during storm events, seasonal cycles align with snowmelt and evapotranspiration dynamics, and spatial transitions correspond to Köppen-Geiger climate zones. Future work may extend this framework to integrate multiple competing physical hypotheses and probabilistic forecasting.