Understanding and Predicting Groundwater Contamination at Nuclear Waste Sites Using Artificial Intelligence and Machine Learning

Abstract:

From the mid to late 20^th century, radioactive waste from the increase in nuclear weapon production during the Cold War was disposed into unlined seepage basins at the F-Area Savannah River Site (SRS) in South Carolina. Many other sites across the United States are of similar concern for groundwater contamination from the migration of acidic plumes due to discharged radioactive waste solutions from this time period. Assessment and remediation at these sites are costly and complex; therefore, user-friendly machine learning tools can help optimize large-scale environmental monitoring across these sites. Focusing on groundwater monitoring data at the Savannah River Site, we have developed a working library of functions in R, with broad applications for other time series data sets. These functions perform a variety of analyses and produce visualizations, to assist in making decisions about monitoring and closure following remediation. Our functions thus far can be used to interpolate time series data, to detect outliers, to plot and smooth trends in concentrations, to perform regression analyses, to predict when analytes will reach their established maximum concentrations limits, to assess correlations between contaminants, and to perform multi-variate regression analyses between related analytes. This R library is also being translated and expanded into a Python package. We plan to further develop this library to create functions for spatial interpolation, multi-temporal scale analysis, and cluster analysis. Through this work, we have discovered patterns and relationships among analytes at this site and made predictions regarding future contaminant concentrations, which will aid in more informed, cost-effective and time-efficient monitoring as these functions are applied across nuclear waste sites.