Deep Physics-Informed Machine Learning Integrating Socio-Economic Indicators for Sustainable Water Governance: A Digital Twin of the Bouregreg Estuary, Morocco

The management of estuarine ecosystem sustainability is a complex problem that requires models that are physically sound, socially meaningful, and interpretable from a mechanistic standpoint. Even though classical AI has demonstrated promise in environmental forecasting, black-box models typically fall short of meeting basic conservation requirements or accounting for anthropogenic stresses that alter water quality. This work introduces a novel framework based on Deep Physics-Informed Neural Networks (Deep PINNs) to predict the dynamics of dissolved oxygen (DO) in the Bouregreg Estuary (Morocco). We advance baseline standards by directly integrating the non-linear advection–diffusion–reaction (ADR) transport equations into the loss function of a deep residual architecture (ResNet with 12–20 layers). This integration ensures that the model takes into account two important aspects of estuarine hydrodynamics: gravitational circulation and the salt wedge effect. The incorporation of a socio–hydro–physical nexus, which uses regional water-pricing indices and urban wastewater discharge volumes from the Rabat-Salé municipal area (∼120,000 m3/day) as proxy variables for anthropogenic pressure, is a unique aspect of this work. The Deep PINN achieves a better coefficient of determination (R2=0.998) and a Nash–Sutcliffe efficiency (NSE=0.997), outperforming the traditional ANFIS and ANN baselines by 89.1% in terms of predictive error reduction (RMSE=0.041±0.002 mg/L). In situations where unconstrained data-driven models fall short, the framework exhibits physical robustness in capturing vertical DO stratification in addition to numerical accuracy. Urban effluent volumes have a significant impact on predictive variance, accounting for 28% of the model internal attribution—more than the relative influence of thermal solubility, according to mechanistic feature attribution analysis using SHAP (Shapley Additive exPlanations). Finally, exploratory management scenarios suggest that summer hypoxia could hypothetically be mitigated through a 20% reduction in discharge volumes. This study bridges the gap between scientific modeling and policy implementation by providing a physics-consistent digital twin framework for environmental stewardship in support of UN SDG 6 and Morocco’s National Water Plan.