SREL Reprint #3709

Artificial intelligence-based predictive modeling for groundwater contamination and physical property forecasting at the Hanford Site - 23271

Michael Murphy1, Hirak Mazumdar1, Hardik Gohel1, Hilary Emerson2, and Daniel Kaplan3

1University Of Houston-Victoria
2Pacific Northwest National Laboratory
3University of Georgia

Abstract: The Department of Energy’s Office of Environmental Management oversees one of the largest groundwater and soil remediation efforts in the world. The Hanford Site, located in Washington State, is home to several contaminated facilities due to previous operations involving nuclear production reactors, laboratories, and large chemical reprocessing plants. While previous research has made considerable progress in characterization and monitoring of groundwater contamination, there are still large plumes of groundwater with contaminants like hexavalent chromium. Recently, machine learning-based methods have shown significant promise in complementing traditional groundwater sampling methods and as decision support systems for developing effective groundwater contaminant management and treatment strategies, thanks to their ability to detect unapparent patterns in complex data. The objective of this work is to review current machine learning techniques for modeling and forecasting groundwater contamination with a focus on (i) regression models due to their effectiveness in dealing with time-series data and (ii) decision tree-based models for their high level of generalizability to a diverse range of problems. The models considered include Multiple Linear Regression, Random Forest, Classification and Regression Trees, and Boosted Regression Trees, and are compared on the bases of their design, the effects of input variability on various performance metrics, overall effectiveness in modeling groundwater features, and methods for improving their performance, such as resampling, imputing missing data, time series feature engineering, hyperparameter tuning, and creating ensembles of multiple models. Then, we implemented a Recurrent-Convolutional Neural Network-based framework for forecasting groundwater features obtained from well measurements taken from six areas on the Hanford Site with hexavalent chromium contamination. The dataset used for model training and testing contains a total of 19,596 measurements collected from 121 wells between 2000 and 2008 for six features including contaminant concentration and five physical properties along with their corresponding timestamps. The model’s performance is evaluated on predicting the concentration of hexavalent chromium and measurements of conductivity, pH, temperature, turbidity, and oxidation-reduction potential using standard metrics such as mean squared error, root mean squared error, and R2 score. The framework presented herein demonstrates that the Recurrent Convolutional Neural Network model outperforms the other models investigated in predicting hexavalent chromium concentration in terms of R2, mean squared error and root mean squared error when trained and tested on the same Hanford Site data. Thus, the methods detailed in this study may prove useful in providing earlier detection of groundwater contaminant changes and aid in the planning and automation of treatment operations.

SREL Reprint #3709

Murphy, M., H. Mazumdar, H. Gohel, H. Emerson, and D. Kaplan. 2023. Artificial intelligence-based predictive modeling for groundwater contamination and physical property forecasting at the Hanford Site - 23271. WM2023. Phoenix, Arizona.

This information was provided by the University of Georgia's Savannah River Ecology Laboratory (srel.uga.edu).