AI-driven groundwater mapping: systematic review and implications for practical uptake

Recent advances in artificial intelligence and machine learning have transformed groundwater mapping by enabling data-driven integration of heterogeneous geospatial, environmental, and hydrogeological information. This paper provides a critical review of AI-based groundwater mapping, synthesizing more than 200 peer-reviewed studies published between 2009 and 2026, with emphasis on the rapid methodological developments of the last 5 years. We outline the conceptual basis of AI-driven mapping approaches as the natural evolution from expert-based GIS overlays and statistical methods. Four dominant areas of application are identified: groundwater potential mapping, spatial prediction of groundwater quality and contamination, vulnerability assessment, and the delineation of groundwater-dependent ecosystems. Ensemble trees, gradient boosting algorithms, and neural networks are the most widely adopted methods, largely due to their ability to capture nonlinear relationships, handle multicollinearity, and perform well with heterogeneous datasets. We also explore the challenges involved in defining the minimum viable dataset size, as well as the increasing interpretability of machine learning outcomes. Despite improvements in predictive accuracy and spatial generalization, this review highlights persistent limitations that constrain real-world adoption. These include limited and uncertain subsurface data, overreliance on surface-derived proxy variables, restricted model transferability, insufficient treatment of uncertainty, and challenges related to model interpretability and transparency. We conclude by identifying priority research directions, including explainable AI, physics-informed and hybrid modeling strategies, spatio-temporal integration, and stronger links between AI outputs and groundwater management decisions.