Do Process-Based Models Really Fall Short? Rethinking Channel Routing to Bridge the Gap with Machine Learning Models

Machine-learning models, particularly Long Short-Term Memory (LSTM) networks, often outperform process-based rainfall–runoff models, yet the specific process limitations driving this performance gap remain underexplored. We hypothesize that a major contributor is the use of simplified channel routing formulations that insufficiently represent temporal variability in flow velocity. Here, we couple the HBV rainfall–runoff model with the recently proposed Iterative Routing Model (IRM), a parsimonious and non-linear channel routing framework that explicitly allows flow velocity to vary with discharge. We evaluate the coupled HBV–IRM model over 64 CAMELS catchments across the United States. The hybrid model attains a median NSE of 0.72, improving on the original HBV (0.65) and approaching the performance of global LSTM benchmarks (0.74). The results indicate that improving process representation in channel routing can substantially reduce the performance gap between process-based and data-driven models, while retaining process understanding and physical interpretability.