Physics-informed spatio-temporal graph neural networks for evapotranspiration prediction: Case of the Korean Peninsula

Study Region: This study focuses on the Korean Peninsula.Study Focus: This study develops a physics-informed spatio-temporal graph neural network for evapotranspiration prediction using climate variables, soil moisture, and a surface energy-balance constraint. The model combines graph-based spatial learning with recurrent temporal learning and is trained using observations from 372 stations from 1950 to 2014. Model skill is evaluated using MAE, RMSE, Nash–Sutcliffe efficiency, and the Continuous Ranked Probability Score. Uncertainty is calibrated using a Monte Carlo dropout and an Isotonic regression. Additional analysis includes feature importance using explainable AI techniques and performance comparison across soil moisture conditions. Climate projections from the Coupled Model Intercomparison Project are used to explore future evapotranspiration responses.New Hydrological Insights for the Region: The model shows strong skill across diverse settings and delivers particularly improved performance under dry conditions, indicating that soil moisture information enhances prediction in water-limited regimes. The energy-balance constraint strengthens physical credibility and improves stability under changing climate forcing. Future climate projections indicate substantial increases in evapotranspiration across much of the Korean Peninsula, with spatial differences linked to temperature, radiation, and moisture availability. These results point to increasing evaporative demand, heightened seasonal water stress, and greater challenges for irrigation and reservoir management. The study demonstrates the value of physics-guided machine learning for regional hydrology and provides decision-relevant insight into future water security under climate change.