Answers are grounded in this paper and cite their sources.
Machine learning is increasingly used to fill gaps in groundwater quality records, yet models are rarely tested in regions they were not trained on. Here nine methods, from logistic regression and raster-based hazard models to a generic tabular foundation model, a groundwater-pretrained Transformer and a simulated quantum kernel, are compared on identical splits drawn from 2.1 million samples. Field parameters and major ions recovered the major-ion chemistry of unseen wells (R² 0.83–0.95), but arsenic reached at most 0.53. Across 47 leave-one-region-out tests, measured chemistry ranked contaminant exceedances with AUC 0.71–0.76, whereas environmental rasters alone stayed near chance (0.58). Out of region, exceedance coverage fell below its nominal 90% for every method, and the best-ranking model missed the most exceedances (35%). Domain pretraining matched a generic tabular model, and the quantum kernel matched its classical counterpart. Ranking accuracy and trustworthy uncertainty therefore separate under geographic shift, and redox-sensitive contaminants still require direct measurement.
No takes yet. Share an insight, caveat, or question.
Hasan et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: