Abstract To address the limitations of process‐driven models in characterizing physical mechanisms and the interpretability challenges of data‐driven models in flood forecasting, this study proposes a distributed hybrid flood modeling (DHFM) framework that integrates physical mechanisms with deep learning. Differentiable diffusion wave (DW) and convolutional neural network (CNN) routing methods are introduced, which can be seamlessly integrated into the DHFM framework. A differentiable Muskingum (MK) routing method is also implemented as a benchmark. The Mishui Basin in China is selected as a case study to systematically evaluate the performance and interpretability of these three routing methods under both gauged and ungauged scenarios. Results show that the DHFM framework can effectively achieve physical parameterization across different sub‐basins. Compared to the lumped Xin'anjiang hydrological model, it achieve s higher accuracy in both daily streamflow and flood simulations, while also demonstrating favorable interpretability of the embedded neural network. Under gauged scenarios, the differentiable CNN method slightly outperforms DW in terms of performance and efficiency, and significantly surpasses MK. As the number of training stations increases, model performance tends to stabilize or decline. In ungauged scenarios, CNN performs well with sufficient training data (>2 stations) but is sensitive to station selection, exhibiting a substantial performance drop with only one station. In contrast, DW and MK show greater stability. The differentiable CNN method shows potential for adaptively learning unit hydrographs based on channel attributes. The proposed DHFM framework not only enhances flood simulation accuracy but also provides novel perspectives for understanding the physical mechanisms underlying flood processes.
He et al. (2025) studied this question.