Machine Learning Spots Uranium in Groundwater From Routine Water Tests

Uranium is one of the most stubborn contaminants that water quality laboratories have to hunt for. It is invisible, tasteless, and chemically mobile in oxygen-rich aquifers, and the only way to know for certain whether a well is safe is to run specialized analytical measurements that are expensive, slow, and impractical to deploy at the scale of thousands of sampling points. A research team based at the Bhabha Atomic Research Centre in Visakhapatnam, working with Andhra University, has now shown that this bottleneck may be far less binding than it appears. In a study published in Environmental Geochemistry and Health, the researchers built a machine learning framework that predicts uranium concentrations in groundwater from ordinary physicochemical parameters that are already measured in routine surveillance programs, potentially transforming how environmental radioactivity monitoring is carried out in vulnerable regions.

The foundation of the work is a decade of patient fieldwork. Between 2016 and 2025, the team collected 1,295 groundwater samples from a coastal region of southeastern India, along the northeastern coast of Andhra Pradesh. Each sample was analyzed not only for uranium but also for a suite of routinely determined parameters such as total dissolved solids, hardness, chloride, and sulphate. The researchers then asked a deceptively simple question: can the cheap, fast measurements alone tell you whether the expensive, slow uranium measurement is likely to come back elevated? If the answer is yes, laboratories could triage their samples, reserving the costly radiometric analyses for the wells that the algorithm flags as high risk, and dramatically cutting the overall screening workload.

Because uranium in drinking water is judged against several different limits rather than one, the team framed the problem as a multi-threshold classification task. They trained their models to distinguish between samples above and below three separate cutoffs: 2 micrograms per liter, which serves as a precautionary level; 15 micrograms per liter, which corresponds to the World Health Organization guideline value; and 30 micrograms per liter, the regulatory limit adopted in India. This design matters because the practical consequences of a prediction differ enormously depending on the threshold. Missing a sample above the regulatory limit is a public health failure, while missing one above the precautionary level is a missed early warning. A monitoring framework that can operate at all three thresholds simultaneously gives regulators a graded picture of risk rather than a single pass-fail verdict.

The modeling itself confronted a classic difficulty in environmental data science: class imbalance. Elevated uranium is, fortunately, a rare event, which means that in any dataset the vast majority of samples fall below the thresholds of interest. Naive classifiers can achieve high apparent accuracy simply by predicting that every sample is safe, while quietly missing the few dangerous ones. To handle this, the researchers evaluated a suite of advanced ensemble learning algorithms, including gradient boosting methods such as CatBoost, LightGBM, and XGBoost, alongside other approaches, and assessed them with metrics like the F1-score that balance precision and recall rather than rewarding the majority class. Hyperparameter optimization was carried out systematically to give each algorithm its best chance of extracting signal from the data.

The results were striking and threshold-dependent. At the precautionary level of 2 micrograms per liter, CatBoost achieved the highest performance, with an F1-score of 80.8 percent, meaning it could reliably flag samples of potential concern from routine chemistry alone. At the WHO guideline level of 15 micrograms per liter, LightGBM performed best, reaching an F1-score of 66.6 percent. The most dramatic result came at the regulatory limit of 30 micrograms per liter, where an Isolation Forest, a one-class anomaly detection algorithm designed specifically for finding rare outliers, achieved 100 percent recall with a false-positive rate of just 1.12 percent. In plain terms, it caught every single sample that exceeded the regulatory limit while wrongly flagging only about one in a hundred safe samples.

That last number translates into an enormous practical saving. Because the anomaly detector essentially never misses a genuine exceedance, laboratories can use it as a first-pass filter: samples the model clears with confidence can skip the specialized uranium analysis, while flagged samples receive full analytical attention. The authors report that this approach reduces laboratory screening workload by 98.7 percent. For a surveillance program that processes hundreds or thousands of samples a year, that figure represents the difference between a monitoring program that is perpetually underfunded and one that can actually cover its territory. It also means that early warning of uranium contamination no longer has to wait for a laboratory queue to clear.

Beyond prediction, the study offers a mechanistic story about why the model works. Using SHAP analysis, a technique from explainable artificial intelligence that quantifies each input variable’s contribution to individual predictions, the researchers found that total dissolved solids, hardness, chloride, and sulphate were the dominant predictors of uranium. These are not arbitrary correlations. Each of them connects to well-established hydrogeochemistry. Total dissolved solids reflect the overall degree of water-rock interaction: the longer groundwater has circulated through the aquifer, the more dissolved minerals, and typically the more uranium, it carries. Hardness, driven by calcium and magnesium, influences the carbonate complexes that keep uranium in solution. Chloride tracks salinity evolution, and sulphate reflects oxidation state and mineral dissolution, both of which govern whether uranium is immobilized or mobilized.

The SHAP findings thus link the model’s statistical behavior to mineral dissolution, salinity evolution, and carbonate complexation, the very processes that geochemists know control uranium mobility. In oxidizing groundwater, uranium readily forms soluble uranyl carbonate species such as UO2(CO3)2(2-) and UO2(CO3)3(4-), which can travel long distances through aquifers without being trapped on mineral surfaces. This is precisely why uranium contamination in India has emerged as a widespread concern in recent years, with large-scale surveys documenting elevated concentrations across multiple states and studies linking the problem to nitrate-driven oxidation and intensive groundwater extraction. The new study adds a coastal dimension to that picture, showing that salinity signals in the water chemistry carry predictive information about radiological risk.

The authors also point toward a complicating future. Climate change may modify the very processes the model exploits, through altered recharge patterns, shifting water-rock interactions, and changing groundwater salinity, particularly in coastal aquifers threatened by seawater intrusion. A machine learning framework trained on the past decade of data will need periodic retraining as hydrological conditions evolve, and the researchers note that their approach is designed to support adaptive monitoring strategies rather than static ones. The framework’s multi-threshold structure is also aligned with the evolving WHO guidelines, which have themselves shifted over the years from an earlier provisional concentration to the current 30 micrograms per liter guideline value, reflecting ongoing reassessment of uranium’s chemical toxicity to the kidney.

What makes this study notable is not any single algorithmic trick but the integration of three ideas that rarely appear together: the use of routine physicochemical predictors that any water laboratory already measures, a multi-threshold classification scheme matched to real regulatory and precautionary levels, and a one-class anomaly detection approach purpose-built for rare-event screening. Together they sketch a blueprint for environmental radioactivity surveillance that is cheaper, faster, and more responsive than the current laboratory-centric model. For the millions of people in India and elsewhere who depend on groundwater for drinking, the practical promise is that the wells most likely to carry a radiological hazard can be identified from data that already exists, before anyone walks into a lab. In a field where the contaminant of concern is invisible and the monitoring budget is finite, that is a genuinely consequential advance.

Subject of Research: Machine learning prediction of uranium concentrations in coastal Indian groundwater from routine physicochemical parameters for environmental radioactivity surveillance

Article Title: Data driven prediction of uranium in groundwater for environmental radioactivity surveillance

Article References: Data driven prediction of uranium in groundwater for environmental radioactivity surveillance. (n.d.). https://doi.org/10.1007/s10653-026-03510-3

Image Credits: AI Generated

DOI: 10.1007/s10653-026-03510-3

Keywords: groundwater, uranium, machine learning, hydrogeochemistry, environmental radioactivity, drinking water, CatBoost, LightGBM, Isolation Forest, SHAP analysis, water quality monitoring, Andhra Pradesh