Physics-informed neural networks for modelling groundwater flow and solute transport in stressed coastal aquifers

Coastal groundwater degradation and salinization from overexploitation and brine mismanagement represent urgent concerns in arid zones. To our knowledge, physics-informed neural networks (PINNs) have not yet been applied to real, highly stressed aquifers affected by seawater intrusion, pumping wells, and discharges from small-scale desalination units. In this study, we present an efficient and environmentally friendly PINN-based framework for simulating groundwater flow and solute transport. The method employs learning rate optimization to simultaneously adjust model parameters and loss function weights. The framework was developed using synthetic datasets generated by MODFLOW and MT3DMS for a coastal aquifer in northern Oman. Performance assessments of hydraulic head and concentration models demonstrated high accuracy across training, validation, and prediction phases. The PINN framework effectively captured complex interactions among recharge, pumping, and boundary conditions, while maintaining strong generalization to unseen data. Parallel coordinate plots and spatial distributions further validated the reliability and convergence behaviour of the model.