Lead Investigator: Haruko M. Wainwright (Massachusetts Institute of Technology)

DOE POC: April Kluever, Charles Denton, Quincy Mason (EM)

Project Objectives:

The goals of this project are to develop a blueprint for archiving/managing/synthesizing groundwater assessment models across multiple sites for enabling cross-site analyses, to automatically detect the deviation of observational data from model predictions, and to facilitate and improve groundwater model developments in the future. Specifically, the project aims to:

• Objective 1: Create the database of past groundwater assessments at the DOE sites such as archiving the model assumptions and parameters as well as the model input files and results (if available) in a machine-readable format.
• Objective 2: Investigate different assumptions and parameter ranges/distributions as well as key factors associated with the success and failures of modeling results.
• Objective 3: Develop a Bayesian model-data assimilation methods and associated AI technologies to automatically ingest datasets collected at the sites and to detect the model deviation/discrepancy from observations.

Significance/Impact:
Many of the groundwater assessments at the DOE sites are known to have failed in the past such that the subsequent observations were not captured within the predicted confidence intervals. Such failures were partly attributed to incomplete scientific understanding (Zachara et al., 2013) as well as to modelers’ judgments. These failures would threaten the credibility of future groundwater assessments as well as performance assessments. At the same time, it is extremely difficult to investigate/analyze past models, since the key information (such as source terms and model parameters) is often buried in large reports and thereby limits independent confirmation.

The database developed in this study will be a valuable asset for DOE-EM for retaining the knowledge relevant to groundwater assessments. In addition, the database will allow us – by taking advantage of artificial intelligence technologies (AI) for pattern recognition – to analyze the commonality and transferability of certain assumptions and parameters. Our results will not only facilitate future groundwater model developments and improve their quality/predictability through the lessons learned. At the same time, the automated model-data assimilation will be able to not only detect anomalies and discrepancies to alert the site managers for re-assessment in a timely manner but also to validate the model results through long-term observations.

Public Benefits:
The improved predictability of groundwater systems using Artificial Intelligence technologies will provide assurance to the public about the performance of waste disposal cells, the stability of residual contaminants, the transparent identification of anomalies, and the valid predictions of groundwater contamination projected forward in time.

References: (* indicates CRESP publication)

Meray, A. O., Sturla, S., Siddiquee, M. R., Serata, R., Uhlemann, S., Gonzalez-Raymat, H., … & Wainwright, H. M. (2022). PyLEnM: A machine learning framework for long-term groundwater contamination monitoring strategies. Environmental science & technology, 56(9), 5973-5983.

*Rustick, J. H., Kosson, D. S., Krahn, S. L., & Clarke, J. H. (2013, July). Building Confidence in LLW Performance Assessments-13386. WM Symposia, 1628 E. Southern Avenue, Suite 9-332, Tempe, AZ 85282 (United States).

Zachara, J. M., Long, P. E., Bargar, J., Davis, J. A., Fox, P., Fredrickson, J. K., … & Yabusaki, S. B. (2013). Persistence of uranium groundwater plumes: Contrasting mechanisms at two DOE sites in the groundwater–river interaction zone. Journal of contaminant hydrology, 147, 45-72.