PINN-Sed: Interconnected Physics-Informed Machine Learning Model for High Resolution Suspended Sediment Transport in River Network

Abstract There is growing demand for river-network models that can estimate sediment transport at fine temporal resolution. Although coupling physically based sediment modules with hydrological and hydrodynamic models is a logical path forward, these approaches require intensive calibration and substantial computing power, limiting their practical use. This study develops a physics-informed neural network (PINN) that embeds sediment transport and advection-dispersion equations directly into a neural-network architecture, boosting predictive accuracy while reducing the need for large observational datasets. We introduce a novel interconnected PINN scheme in which PINNs for separate river reaches are coupled through loss functions at their boundaries, creating a cohesive model for the river network. The resulting framework predicts suspended-sediment concentrations reliably throughout the catchment during flood events. We evaluated three configurations of the PINN model and compared their performance against a purely data-driven machine learning model and an optimized physically based model. Trained on a subset of flood events and tested on unseen events, the interconnected PINN framework (root mean squared error = 5.8 kt) outperformed both the data-driven (33.8 kt) and physical (27.8 kt) models in estimating event loads. This novel PINN framework is transferable and can be applied to other catchment and river-network modeling, supporting simulations of both hydrology and flow-transported constituents. It is among the first to leverage comprehensive real-world data, using observations from the Manawatū River in New Zealand for both model training and validation. Requiring only standard river-network descriptors and hydrological time series, the approach can be transferred readily to other catchments with similar datasets.