Application of a hybrid GR4J-PINNS framework for enhanced extreme flood peak simulation in a tropical Monsoon Basin, Vietnam

Conceptual rainfall–runoff models such as GR4J provide a parsimonious and physically interpretable framework for streamflow simulation, but their accuracy may decline during extreme floods in tropical monsoon catchments with concentrated rainfall and nonlinear runoff responses. This study presents a hybrid GR4J–Physics-Informed Neural Network (GR4J–PINNs) framework to improve flood-peak simulation while preserving the physical interpretability of GR4J. The calibrated GR4J model is treated as a frozen physical trajectory, while a neural network learns the residual error between simulated and observed discharge under a physics-consistency constraint that penalizes departures from the mass-conservative GR4J solution. To address data scarcity in the Vu Gia–Thu Bon River Basin, Vietnam, the framework incorporates NASA GLDAS satellite-derived evapotranspiration and a regression-based baseflow index derived from Lyne–Hollick filtering without using observed discharge as an input. Using a chronological train/test partition (2010–2020) and a prospective application trial (2021–2022), GR4J–PINNs improves NSE from 0.882 to 0.967 and R2 from 0.939 to 0.983, while reducing MAE and RMSE by 61.8 and 46.7%, respectively. The results demonstrate improved flood-peak simulation and robust performance, highlighting GR4J–PINNs as a non-intrusive, auditable approach for integrating conceptual hydrological models with deep learning in data-scarce tropical basins.