Evaluating physics informed neural networks for water contamination risk prediction and environmental sustainability

Growing pollution hazards in large watersheds, including Egypt's Nile River, necessitate precise and effective monitoring of water quality indices. Although considerable efforts have been made to model individual parameters such as pH, Total Dissolved Solids (TDS), Electrical Conductivity (EC) and sodium (Na[Formula: see text]), the emphasis on isolated forecasts has often overlooked their interdependencies, thereby limiting estimation accuracy and practical applicability. Conventional machine learning methodologies further compound this issue, as they frequently lack adaptability and physical interpretability, resulting in inefficiencies for sustainable water management. To address these limitations, this study develops a physics-informed neural network (PINN) framework integrated with optimization boosting techniques to jointly predict pH, TDS (mg/L), EC (μS/cm), and Na[Formula: see text] (mg/L) under three critical management strategies: industrial discharge regulation, salinity management and irrigation planning. The approach incorporates prior hydrodynamic and chemical knowledge to guide input categorization, employs adaptive weighting mechanisms to dynamically adjust feature relevance, and introduces a deep interaction module to capture intricate physicochemical couplings. A physics-constrained loss function ensures consistency with ecological processes. Field analysis revealed that while pH values (6.01-6.87) consistently met FAO standards, TDS, EC, and Na[Formula: see text] frequently exceeded permissible thresholds, particularly during the dry season. Model evaluation showed that PINN outperformed conventional optimizers, achieving [Formula: see text] values of 0.945-0.999 and RMSE between 0.012-0.088. Cross-validation further confirmed its robustness, yielding consistently low RMSE and near-unity [Formula: see text] across folds. Interpretability analysis highlighted irrigation intensity, salinity loads, and industrial effluents as the dominant drivers of water quality dynamics. Overall, the proposed PINN not only advances predictive accuracy and computational efficiency but also provides a practical framework for supporting regulatory decision-making, optimizing salinity control, and guiding sustainable irrigation planning. Ultimately, this approach offers a viable pathway for mitigating contamination risks and ensuring the long-term sustainability of the Nile River ecosystem.