A model of groundwater in Colorado's San Luis Valley assigned higher probabilities of high arsenic to deeper wells, while direct measurements found arsenic at or above 5 micrograms per liter in 19.8% of private wells.
A wide range of wells and readings
The study combined community well samples with well records. Community sampling between 2022 and 2024 produced 613 wells; 459 had boron and lithium measurements, and 435 had complete predictor data and entered the random-forest analysis. A separate time-based analysis used records for 16,221 domestic or household wells drilled since 1970. Those wells did not have direct arsenic measurements.
To make the prediction, the researchers trained a random-forest classifier to sort wells into two groups: those at or above the 5 micrograms per liter cutoff and those below it. They used a random 70% of the complete wells for training, then evaluated the model on held-out data and checked its settings through 10-fold cross-validation repeated five times. They compared a baseline model with SMOTE, a resampling approach for class imbalance.
The reported measurements covered a wide range. In the 435 modeled wells, arsenic was nondetectable in 22 samples and reached 114.65 micrograms per liter; the median was 1.27. Well depths ranged from 5.5 to 938.8 meters, with a median of 32.0 meters.
The model was tested on unseen data
On held-out testing data, the baseline model had an AUC, a measure of how well it separated the two groups, of 0.879. Its specificity was 0.781 and its sensitivity was 0.885. The SMOTE version was more specific, at 0.819, and had a sensitivity of 0.808, with an AUC of 0.873.
Depth was one of the clearest signals
Well depth, elevation, temperature, soil organic carbon and geothermal activity were among the strongest signals in the predictor rankings. The model assigned higher arsenic probabilities to deeper wells, lower elevations, greater geothermal activity, more fine-grained sediment and more soil clay. These are model-based associations, not evidence that any one factor causes arsenic levels to rise.
Long-term subsidence, which the study considered as a pumping-related signal, had only a weak positive relationship with arsenic and was less influential than source-water and natural geological conditions. The model therefore gave greater weight to those conditions than to subsidence.
The researchers also produced what they called a pseudo-three-dimensional map, a set of predictions at selected well depths. Maps were generated at 25, 50 and 75 meters because 83% of study wells were shallower than 75 meters. The approach still used depth as a predictor but did not account for depth-wise changes in lithology or geochemical conditions.
County trends did not move together
To examine change over time, the final model classified the 16,221 domestic or household wells at the same 5 micrograms per liter cutoff, even though the records contained no direct arsenic measurements. The county patterns diverged. Since 1970, Alamosa showed a slight decrease in modeled arsenic and well depth, while Conejos and Rio Grande showed increasing well depth and predicted arsenic exposure. The reasons private-well depths changed differently among counties were not known.
That distinction matters: the county trends are model estimates, not measurements from those 16,221 wells. The study's direct measurements came from community-sampled wells, and the researchers did not make measurement locations or the underlying data publicly available because of participant privacy.
A regional prediction, not a causal answer
Taken together, the results describe regional probabilities and associations. They do not establish that depth, geothermal activity, subsidence or any other predictor causes arsenic enrichment, and the temporal analysis cannot substitute for direct arsenic measurements in the wells it modeled.
The study reports funding from the National Institute of Environmental Health Sciences under award R01ES032612.
Paper data and sources
Original title: Anthropogenic and Geogenic Factors Affecting Groundwater Arsenic and Potential Drinking Water Exposure in the San Luis Valley, Colorado
Authors: Smith R, Lopez A, Honeyman A et al.
Journal/Repository: Environmental science & technology
Status: Peer-reviewed
First online: 2026-10-19
DOI: Not available
Original paper · Full text