Design and Application of a Physics-Informed Machine Learning Model for Salinity Prediction in the Nakdong River Estuary

This paper proposes a hybrid framework to predict salinity in the Nakdong River Estuary by combining physics-based modeling with machine learning. Motivated by the increasing complexity in salinity intrusion owing to frequent barrage operations and nonlinear hydrological changes, we developed a physics-informed machine learning (PIML) model. The model integrates a physical salinity intrusion curve with a long short-term memory (LSTM) network, using a customized loss function that dynamically balances each component. Validation was performed using observation data from January to June 2022 at multiple hydrological stations. A moving window method was applied for sequential training, and a genetic algorithm was used to optimize physical parameters. The results showed that during sharp salinity surges, the physical model provided more accurate predictions, whereas the data-driven model performed better during stable periods. Overall, the PIML approach achieved the lowest mean squared error (MSE) and demonstrated robust and generalized performance. These results highlight the model’s ability to enhance prediction reliability by selectively emphasizing physical or data-driven components, which is particularly useful in estuarine environments with variable tidal and inflow dynamics.