PREDICTING GROUNDWATER QUALITY IN DRINKING WATER WELLS IN THE GLACIAL AQUIFER SYSTEM, NORTHERN USA

Chronic exposure to geogenic contaminants (e.g., arsenic (As) and manganese (Mn)) or anthropogenic contaminants (e.g., nitrate (NO₃)) via drinking water is a human health concern worldwide. The glacial aquifer system in the northern contiguous United States overlies portions of 24 states and ranks first in the Nation as a source of groundwater for both public and domestic water supply. Across its extent from Maine to Washington, sediments of glacial origin vary in thickness, texture and water-quality conditions. Environmental factors such as climate, precipitation, and soil chemistry and minerology are highly variable across the region, too.

The solubility, mobility, and degradation of groundwater contaminants are dependent on many geochemical factors. Redox conditions and pH are particularly important geochemical factors because they affect the solubility and mobility of contaminants such as As, Mn and NO₃ through processes including degradation, ion exchange, sorption, complexation, denitrification and mineral saturation. Machine learning methods are being used to develop aquifer-wide, 3-D models of pH and redox conditions using water quality data available from the USGS National Water Information System database, State ambient and compliance monitoring data, and predictor variables from numerous sources.

Model predictor variables – including soil chemistry, land use, aquifer texture and hydrology, hydrologic position, well construction information and groundwater age – are compiled into training and hold-out data sets, respectively, and applied to Boosted Regression Tree (BRT) prediction models. The relative contributions of the various explanatory variables are evaluated. Partial dependency plots can be used to infer the direction and degree of influence that explanatory factors have on the response variables. Predictor variables are proxies for measured physical and geochemical processes affecting contaminant mobilization. Grids of continuous predicted pH and redox ranges for drinking water depth intervals are generated by applying the final BRT model to gridded predictor variables. The gridded outputs of the redox and pH models will then be used as inputs into models to predict occurrence of contaminants that affect drinking water quality, such as As, Mn, or NO₃.