Machine Learning-Driven Framework for Groundwater Quality Monitoring in Arid Coastal Aquifers

The strategic design of groundwater quality surveillance systems (GQSS) is essential for accurate monitoring and timely detection of groundwater contamination. AI-assisted optimization approaches, particularly the Ant System meta-heuristic algorithm (AS-MHA), are effective for complex spatial allocation problems, but conventional methods rely on fixed parameters, limiting global search efficiency and slowing convergence. To overcome these limitations, a Dynamic Parameter Optimization Framework based on the Ant System (DPOF-AS) was developed for adaptive parameter tuning. The algorithm was applied to an agriculturally intensive aquifer in Mareth, Southern Tunisia, calibrated with continuous groundwater-level observations and 290 water quality samples collected from wells and boreholes between February and May 2025, and validated against an independent dataset spanning 7 May to 5 August 2025. A Support Vector Regression enhanced with Genetic Algorithm (SVR-GA) was employed as a surrogate model, reducing computational cost and accounting for uncertainties related to aquifer heterogeneity and pollution sources. The surrogate model achieved R² of 0.97 and MRE of 4.6% for the flow model, R² of 0.95 and MRE of 6.8% for nitrate transport, and R² > 0.94 with MRE < 7% for SVR-GA predictions, confirming high predictive accuracy. The DPOF-AS framework generated three optimized monitoring layouts (63, 28, and 15 wells for 1, 2, and 5 km surveillance radii), maintaining spatial representativeness, reducing redundancy, and achieving an average pollutant detection rate of 93.5%, outperforming conventional and random placement strategies. This study demonstrates the novelty of integrating adaptive parameter optimization with AI-enhanced surrogate modeling, providing a practical, cost-effective tool for designing reliable and timely groundwater pollution monitoring networks in complex aquifer systems.