Improving large-basin streamflow simulation using a modular, differentiable, learnable graph model for routing

Differentiable modeling has been introduced recently as a method to learn relationships from a combination of data and structural priors. This method uses end-to-end gradient tracking inside a process-based model to tune internal states and parameters along with neural networks, allowing us to learn underlying processes and spatial patterns. Hydrologic routing modules are typically needed to simulate flows in stem rivers downstream of large, heterogeneous basins, but obtaining suitable parameterization for them has previously been difficult. In this work, we apply differentiable modeling in the scope of streamflow prediction by coupling a physically-based routing model (which computes flow velocity and discharge in the river network given upstream inflow conditions) to neural networks which provide parameterizations for Manning’s river roughness parameter (n). This method consists of an embedded Neural Network (NN), which uses (imperfect) DL-simulated runoffs and reach-scale attributes as forcings and inputs, respectively, entered into the Muskingum-Cunge method and trained solely on downstream discharge. Our initial results show that while we cannot identify channel geometries, we can learn a parameterization scheme for roughness that follows observed n trends. Training on a short sample of observed data showed that we could obtain highly accurate routing results for the training and inner, untrained gages. This general framework can be applied to small and large scales to learn channel roughness and predict streamflow with heightened interpretability.