Modeling and mapping high arsenic and manganese concentrations in the glacial aquifer system, northern USA: Using machine learning methods to predict water quality conditions in an extensive, stratigraphically complex, unconsolidated drinking water aquifer system

The glacial aquifer system in the northern continental United States overlies portions of 24 states and supplies drinking water to about 30 million people - more people than any other aquifer in the country. Across its extent from Maine to Washington, sediments of Pleistocene age vary in thickness and water-quality conditions; high concentrations of geogenic As (>10 µg/L) and Mn (>300 µg/L) are widespread. Chronic exposures from drinking water is a human-health concern because of (1) As: increased risks for certain cancers, skin abnormalities, and peripheral neuropathy; (2) Mn: brain and neurological damage, especially in infants. Understanding occurrence and concentrations of As and Mn is, therefore, vital to reducing exposure from drinking water.

Across the glacial aquifer system, occurrence and concentration of As and Mn vary considerably because of differences in climate, soil/aquifer chemistry, geochemical processes, hydrologic position, and groundwater residence time. Because of the complexity and scale of the glacial aquifer system, machine learning modeling methods have better predictive power compared to standard geochemical or statistical methods.

Boosted regression tree models are being used to predict the probability of high As and high Mn, evaluate the relative importance of predictor variables, and describe relations between predictor variables and high As or Mn concentrations. Model inputs include data from numerous sources: As and Mn water-quality data, soil chemistry, aquifer texture and thickness, hydrologic position, well and water table depth, groundwater age, etc. The use of new 3D models of pH and redox condition as predictor variables is novel and particularly important because redox and pH affect the solubility/mobility of As and Mn. Our methods are transferable to other settings or constituents.

Influential predictors of high As include likelihood of anoxic groundwater conditions, concentrations of As in soil, and pH. The most influential predictors of high Mn include recharge, hydrologic position, well depth, and pH. Preliminary results (attached figure) demonstrate that these new tools help identify areas at greater risk for exceedances of health standards and better inform drinking water regulators, public health professionals, water suppliers and well owners.