Predicting and managing seawater intrusion using a hybrid machine learning–reinforcement learning framework: The Mnasra coastal aquifer, Northwestern Morocco

Coastal aquifers are essential freshwater sources for irrigated agriculture in semi-arid regions, but they are increasingly threatened by seawater intrusion driven by intensive pumping and climate-related stressors. This study develops a preliminary hybrid machine learning–reinforcement learning (ML–RL) framework to explore seawater-intrusion management scenarios in the Mnasra coastal aquifer, northwestern Morocco. The Seawater Intrusion Index (SWI), defined as the seawater fraction, was analytically derived from chloride concentration. Ridge Regression was used as a static surrogate estimator of this chloride-derived SWI using 13 hydrochemical variables measured in 36 wells from a single hydrochemical sampling campaign. Because chloride is directly involved in the analytical definition of SWI, predictive performance was interpreted cautiously. Model robustness was assessed using chloride-ablation analysis, leave-one-out cross-validation, repeated 5-fold cross-validation, and feature-contribution diagnostics. The full model including Cl⁻ reproduced the chloride-derived SWI with high cross-validated performance, whereas the chloride-excluded model provided a more conservative assessment of the broader hydrochemical signature. In the second stage, a Deep Q-Network (DQN) agent explored well-specific pumping-reduction actions (0%, 19%, 30%, and 40%) within a simplified synthetic surrogate environment initialized from the observed hydrochemical state. The DQN component was designed to reduce SWI toward the agricultural protection threshold of 0.01 while preserving usable groundwater supply. Because the dataset contains no observed temporal pumping-response trajectories, the RL outputs are interpreted as synthetic scenario-screening results rather than field-validated dynamic predictions. The management scenarios indicate partial mitigation of salinization: SWI was reduced in moderate- and high-intrusion wells while a substantial fraction of water supply was preserved, but several wells remained above the 0.01 threshold after management. The proposed framework should therefore be viewed as a proof-of-concept decision-support approach that complements, rather than replaces, geological characterization and process-based density-dependent groundwater models such as SEAWAT or FEFLOW.