A machine learning approach for prioritizing groundwater testing for per-and polyfluoroalkyl substances (PFAS)

Area of Science:

• Environmental Science
• Hydrogeology
• Data Science

Background:

• Per- and polyfluoroalkyl substances (PFAS) are recognized as significant environmental contaminants.
• Groundwater testing and mitigation for PFAS are costly and time-intensive, leading to limited well coverage.
• This results in potential prolonged exposure to PFAS in drinking water.

Purpose of the Study:

• To develop and compare machine learning models for predicting PFAS concentrations in groundwater.
• To assess the predictive power of various data types, including co-contaminants, hydrology, soil, proximity to sources, and geospatial data.
• To establish a practical tool for identifying wells with high PFAS levels.

Main Methods:

• Construction of multiple machine learning models, including linear and random forest regressors.
• Utilizing a comprehensive groundwater dataset from California.
• Evaluating the predictive performance of different feature sets and a combined model.

Main Results:

• A random forest model combining all data types achieved a Spearman correlation of 0.64 for quantitative PFAS concentration prediction.
• The model accurately identified wells with concerningly high PFAS concentrations with 91% accuracy (AUC of 0.90).
• Co-contaminant fingerprints, hydrological properties, soil parameters, proximity to airports/military bases, and geospatial data were assessed for predictive ability.

Conclusions:

• Machine learning offers a viable approach for predicting PFAS contamination in groundwater.
• This method can efficiently identify high-risk wells, potentially prioritizing testing and mitigation efforts.
• The approach shows promise for application to other hazardous anthropogenic groundwater contaminants.