Unraveling Hydrochemical Drivers of Groundwater Quality and Assessing Associated Health Risks Using Self-Organizing Map, Explainable Artificial Intelligence, and Monte Carlo Simulation

The desert steppe on the northern foothills of the Yinshan Mountains serves as a critical ecological barrier and an agro-pastoral ecotone in northern China. Sustainable utilization of groundwater resources is essential for safeguarding regional ecological security and supporting socio-economic development. In the region, groundwater not only sustains fragile ecosystems but also constitutes the sole source of drinking water for local. However, the groundwater quality exhibits spatial heterogeneity due to the complex geological settings and anthropogenic activities. The study focused on the Tabu River Basin, a representative area of the desert steppe, where 107 groundwater samples were collected. By integrating conventional hydrochemical analysis, self-organizing map (SOM), explainable artificial intelligence (XAI) methods, and health risk assessment coupled with Monte Carlo simulation, we systematically characterized groundwater chemistry, evaluated its suitability for drinking purposes, and identified the dominant factors controlling water quality variations. The results showed that the self-organizing map classified three groundwater clusters, and hydrochemical facies were primarily identified as HCO₃⁻–Ca²⁺, Cl⁻–Na⁺, and mixed HCO₃⁻–Ca²⁺·Na⁺ types, primarily governed by cation exchange and human activities.  The entropy-weighted water quality index (EWQI) showed that 23.3% of the samples were classified as excellent, 36.5% as moderate, while 15.9% and 24.3% fell into the poor and very poor categories, respectively. Further analysis employing the XGBoost model combined with SHAP (Shapley Additive Explanations) interpretability techniques identified nitrate (NO₃⁻) and total dissolved solids (TDS) as the key drivers of water quality deterioration. Health risk assessment results indicated that 98.9%, 92.0%, and 80.5% of groundwater samples exceeded the acceptable threshold for total non-carcinogenic health risks for children, adult females, and adult males, respectively. By synergistically combining traditional hydrochemical approaches with unsupervised machine learning (SOM) and interpretable machine learning (XGBoost+SHAP), the study establishes a multidimensional and highly interpretable analytical framework, which not only advances the understanding of groundwater evolution mechanisms in arid and semi-arid inland basins but also provides robust scientific support for the sustainable management and utilization of regional groundwater resources.