Integrating hydraulic processes and graph neural networks for event-scale river morphodynamics

Abstract

Riverbank erosion is a fundamental process shaping channel morphology, sediment dynamics, and floodplain evolution, yet its prediction remains challenging due to the nonlinear interplay between hydraulic forcing, sediment transport, and geotechnical resistance. Traditional physics-based models grounded in the shallow water equations and the Exner sediment continuity equation provide mechanistic insights but are computationally intensive and sensitive to parameter uncertainties. Conversely, machine learning methods offer rapid and accurate predictions but often lack physical interpretability and generalization. This study introduces a novel spatio-temporal graph neural network (ST-GNN) with a GRU backbone, designed to capture directed hydraulic connectivity and dynamic morphodynamic interactions across a river network. The framework of this research is developed as a surrogate model trained on high-fidelity labels derived from a validated hydromorphodynamic model that explicitly incorporates toe scour, excess-slope failure, and toe-transport mechanisms. This approach enables the ST-GNN to emulate complex morphodynamic responses with high computational efficiency and accuracy, providing a reliable approximation of process-based simulations. The key innovation of this work lies in leveraging a physics-guided graph topology to represent hydraulic connectivity within a machine learning framework. Results demonstrate that the model effectively captures reach-consistent spatio-temporal patterns of bed and bank adjustments. This approach contributes to the predictive science of river morphodynamics by establishing a high-fidelity surrogate that maintains physical consistency while overcoming the computational bottlenecks of traditional numerical solvers.