Long Short-Term Memory Networks for Monitoring Groundwater Contamination at the Hanford Site

The U. S. Department of Energy's Office of Environmental Management handles one of the world's most significant groundwater and soil remediation efforts. The Hanford Site in Washington State contains several decommissioned nuclear production reactors, laboratories, and chemical reprocessing plants, which are the source of various contaminants of concern in groundwater reservoirs. While previous research has yielded significant insight into the behavior of groundwater contaminants at the Site, plumes that contain carcinogens such as hexavalent chromium can be challenging to model using traditional physical and statistical models. Recently, machine learning models, and specifically artificial neural network models, have shown promise in complementing existing methods due to their effectiveness in modeling sequential data, adaptability to different datatypes, and non-fixed parameters with the ability for fine-tuning. In this study, we propose a Long Short-Term Memory-based framework for predicting hexavalent chromium concentration at the Hanford Site using a dataset containing 2912 measurements collected from 121 wells between 2000 and 2008. Both the unoptimized and optimized versions of the model are evaluated using standard metrics including mean squared error, root mean squared error, and coefficient of determination. The optimized model achieves a mean squared error of 921.0, root mean squared error of 30.35, and coefficient of determination score of 0.94, indicating that the use of such network architecture may prove useful in aiding existing modeling operations for contaminants of concern at the Hanford Site and other similar facilities.