Optimized Machine Learning Model for Predicting Groundwater Contamination

The use of physical models to predict groundwater contaminant movement remains technically challenging due to the complexity of the phenomena, the heterogeneity of key parameters in nature, and the presence of poorly defined interactive and feedback processes. New approaches to address these challenges are needed. In this study, we evaluate various Artificial Intelligence (AI)-based approaches to understand hexavalent chromium (Cr(VI)) plumes located on the U.S. Department of Energy’s (DOE) Hanford Site in Richland, WA. The groundwater monitoring dataset used in this study included data from the 100 Area along the Columbia River and included data collected between 2010 to 2019. This study investigates the most prominent contaminant, Cr(VI), with the Extreme Gradient Boosting (XGBoost) machine learning model. The XGBoost model was compared with an optimized version using an Empirical Bayes Search Cross-Validation technique for better prediction. The optimized XGBoost model yielded an R2 value of 0.99 on the training set and 0.85 on the testing set, whereas X G B Boost without optimization yielded a value of 0.83 on the training set and 0.73 on the testing set. This paper provides an overview of a computational method for groundwater contamination modeling that shows promise for improving current remediation efforts.