Opening Remarks
 

1:05 - 1:25 p.m.

Simulator-Trained AI for Creating Subsurface Digital Twins using Time-Lapse Electrical Resistivity Tomography Data

Tim C. Johnson, Pacific Northwest National Laboratory

Electrical Resistivity Tomography (ERT) has been used extensively to monitor for spatial and temporal changes in subsurface bulk electrical conductivity associated with subsurface remediation processes. Changes in resistivity are implicitly governed by hydrogeologic parameters that determine remediation performance.  For example, the spatial distribution of permeability influences the migration and distribution of injected remedial amendments. ERT measurements that sense the amendment migration are therefore sensitive to the permeability distribution. Likewise, the downward migration of flush water through the vadose zone is governed by the moisture retention properties of vadose zone soils. Therefore, ERT measurements that sense the changes in saturation over time during in-situ soil flushing are sensitive to the moisture retention curves that govern how flush water migrates through the vadose zone.

Numerical simulators play a ubiquitous role in subsurface remediation by incorporating contemporary knowledge of subsurface physical and chemical processes through the equations by which they’re described, aspirationally constituting a physics based digital twin of the subsurface. It is well known that the primary source of uncertainty in subsurface numerical model predictions is the corresponding uncertainty in the heterogeneous hydrogeologic properties that serve as model inputs. Time-lapse ERT data, in addition to traditional monitoring data, potentially provide a rich source of information regarding hydrogeologic properties that can reduce uncertainty and improve that accuracy of numerical models that are used in subsurface management.

In this talk, we demonstrate how the hydrogeologic information content implicit in time-lapse ERT data can be extracted and assessed using generative artificial intelligence (AI).  The trained AI model generates ensembles of subsurface models that honor field observations, thereby providing a mechanism for assessing uncertainty.  We demonstrate the approach on the Hanford 100K-Area situ soil flushing test, where ERT was used to monitor vadose zone hexavalent chromium remediation through in situ soil flushing.

Coauthors: Jose Hernandez-Mejia, Glenn Hammond, Piyoosh Jaysaval, and Rob Mackley (PNNL)

1:25 - 1:45 p.m.

Analysis of Chernobyl Groundwater Monitoring Data Using Unsupervised and Supervised Machine Learning Algorithms

Dmitri Bugai, Ukraine Institute of Geological Sciences

A pilot analysis using unsupervised machine learning was conducted to identify factors potentially influencing 90Sr and 137Cs concentrations in groundwater in the Chornobyl Exclusion Zone. The dataset included measurements from 72 monitoring wells collected between 1989 and 2022. The analysis employed the PyLEnM software (Python for Long-term Environmental Monitoring), developed at the Pacific Northwest National Laboratory.

No strong positive correlation was found between radionuclide concentrations in groundwater and land surface contamination. The limited set of variables used in the initial analysis—surface contamination density, depth to groundwater, and well screen depth—appeared insufficient to explain the variability of observed radionuclide concentrations. A negative correlation between 90Sr concentrations and sampling depth supports a conceptual model involving migration from a near-surface source. In contrast, correlation patterns for 137Cs do not support advective-dispersive transport from a near-surface source and may indicate borehole contamination during drilling.

In the second phase of the study, supervised learning techniques were applied to explore the potential for predicting ^90Sr concentrations using an expanded set of features. These included categorical variables such as source type (contaminated topsoil vs. waste dumps), sampling position within the aquifer (top vs. bottom), and screen length (short vs. long, indicating potential for sample dilution). A Random Forest model trained on these features achieved moderate prediction accuracy, highlighting the importance of both source characteristics and hydrogeological parameters (e.g., depth to groundwater, sampling depth) in understanding 90Sr variability.

Coauthor: Dmitri Grygorenko (Ukraine Institute of Geological Sciences)

1:45 - 2:05 p.m.

Characterisation of Hydrogeological Systems using Timeseries Analysis

Joel Wells, Sellafield, Ltd.

On remediation sites with complex or highly heterogeneous hydrogeological systems, obtaining a detailed understanding of the groundwater system behaviour is often of crucial importance to the success of mitigation measures. In this talk, we explore how several time-series methods can analyse short and longer-term temporal changes in groundwater heads across a complex site, delivering new insights into system behaviour.

Firstly, a spectral (Fourier) analysis method is presented to analyse short-term harmonic tidal signals observed in high-frequency level data collected from in-situ level sensors. This analysis separates and quantifies the amplitudes and phases of harmonics at different frequencies, quantifying the effect of earth-tide and oceanic-tide forcing on the groundwater system. This approach concludes with the estimation of hydrogeological parameters using analytical methods, aiding in complex model calibration.

Secondly, an analysis of longer-term level measurements is presented; seasonal ARIMA models are explored to estimate monthly-scale evolution of groundwater levels and relationship to long-term rainfall data. This approach is developed to enable forward-forecasting of level data across site and optimisation of monitoring to areas where level variations are less well understood. The analysis also provides an ‘early-warning’ system for changes in system behaviour over time, such as through climate change or changing leak behaviour.

Finally, the first two approaches are combined to develop a high-temporal-resolution forecasting tool using modern machine-learning methods for time-series. Throughout the talk, the challenges and benefits of each approach will be discussed, alongside presentation of a practical case study using data from the Sellafield nuclear site in Cumbria, UK.

2:05 - 2:25 p.m.

ALTEMIS: Next-Generation In-Situ Real-time Groundwater Monitoring Strategies

Haruko Wainwright, Massachusetts Institute of Technology

The Advanced Long-Term Monitoring Systems (ALTEMIS) project is developing an innovative paradigm of long-term monitoring based on state-of-art technologies – in situ groundwater sensors, geophysics, drone/satellite-based remote sensing, reactive transport modeling, and AI – that will improve effectiveness and robustness, while reducing the overall cost. As a part of this project, this study aims to develop an in situ real-time groundwater long-term monitoring (LTM) framework based on various sensors and data analytics methods. Rather than relying on one single metric, our approach provides multiple lines of evidence to ensure the system stability: (1) groundwater table and its gradient that governs the migration speed and direction of the contaminant plume, and (2) in situ measurable geochemical parameters (specific conductance, pH, and others) for detecting changes in contaminant mobility. In addition, we develop machine learning algorithms to (a) improve the spatiotemporal interpolation of groundwater tables and contaminant concentrations by exploiting proxy variables such as in situ sensors and geospatial layers, and (b) detect anomalies by computing the difference between near-future forecasting and measurements. To accommodate noisy and drifting sensor data streams, we also developed algorithms for automated outlier removal and drift correction. We demonstrate the framework based on the plot-scale installation at the Savannah River Site F-Area in the US.

Coauthors: Haokai Zhao (MIT); Hansell Gonzalez-Raymat, Thomas Danielson, and Carol Eddy-Dilek (Savannah River National Laboratory)

Open Discussion
 

BREAK
 

3:15 - 3:35 p.m.

Forecasting Groundwater Levels and Optimizing Monitoring Networks for Remediation Design Using Diffusion Models

Xuehang Song, Pacific Northwest National Laboratory

Effective environmental remediation at complex sites like Hanford relies on accurate forecasting of groundwater levels to support long-term performance monitoring and decision-making. Groundwater levels (i.e., hydraulic heads) are foundational data for designing and evaluating groundwater remedies—informing flow models, estimating gradients, and guiding system operations.

We present an AI-based framework that combines denoising diffusion probabilistic models with score-based data assimilation (SDA) to forecast groundwater level evolution and optimize monitoring well networks. The framework is trained on high-resolution, physics-based simulations from a calibrated groundwater flow model spanning over 75 years of monthly time steps. The unconditional diffusion model treats groundwater levels as spatial-temporal “videos” across a structured grid, capturing long-term trends and variability. During inference, SDA enables assimilation of sparse and irregular observations, ensuring predictions remain consistent with real-world data without retraining. This supports flexible integration of new measurements and robust forecasting under varying data availability.

Preliminary results demonstrate strong performance in both forecasting and data gap-filling tasks. The model accurately reconstructs spatial patterns and temporal evolution of groundwater levels, even under limited observation coverage. To support remediation monitoring network design, the framework quantifies predictive uncertainty and evaluates each well’s contribution to reducing it. This informs optimal placement of new wells and identification of redundant or low-value ones. Overall, this work highlights how generative AI can enhance groundwater level forecasting, reduce reliance on dense sensor arrays, and inform scalable, adaptive monitoring strategies for remediation. The approach is transferable to other complex sites and aligns with the growing need for cost-effective, data-driven environmental management solutions.

Coauthors: Tse-Chun Chen, Grigoriy Kondyukov, Zhangshuan Hou, Inci Demirkanli, and Rob Mackley (PNNL)

3:35 - 3:55 p.m.

AI-Powered Waste Intelligence: Enhancing Environmental Remediation Through Autonomous Robotics and Data Fusion

Halina Harvey, Innovative Physics, Ltd.

Advanced AI technologies are reshaping environmental remediation. At the heart of this transformation is the Intelligent Robotic Operating System (IROS), developed by Innovative Physics, which autonomously sorts, classifies, and segregates complex nuclear waste materials using advanced computer vision, deep learning, and multi-sensor data fusion.

IROS leverages machine learning models trained on orthogonal datasets—including radiation signatures, density, ferromagnetism, and photogrammetry—to deliver real-time, high-confidence identification of waste materials and their contamination levels. Through a custom SQL-based decision engine, the system dynamically applies Waste Acceptance Criteria (WAC), adapting instantly to site-specific remediation rules and updating logic through modular configuration rather than code.

Environmental remediation efforts benefit significantly from IROS’s capability to process vast and complex data from multi-modal sensors (radiation, LiDAR, multispectral, 3D SLAM). The AI not only guides item-level classification but also plans grasping strategies in cluttered, unstructured waste environments. Real-time pose estimation and 3D mapping drive safe manipulation, while formal software verification ensures every robotic action is inherently safe and traceable.

By removing humans from direct exposure, improving classification accuracy, and accelerating operational throughput, IROS reduces the environmental impact of misclassified waste and contributes to faster site recovery. Moreover, the system stores operational data for long-term analytics, enabling smarter resource deployment and adaptive learning for future clean-up campaigns.

This platform demonstrates how AI is more than an assistive tool—it is a central agent in intelligent remediation. IROS embodies a scalable and transferable solution that integrates robotics, analytics, and AI into a unified system for cleaner, faster, and safer environmental outcomes.

Coauthors: Mike Anderson and David Prendergast (Innovative Physics, Ltd.)
 

3:55 - 4:15 p.m.

From Data to Decisions: AI-Driven Innovation in Environmental Remediation with EQuIS 

Taylor Ziolkowski, EarthSoft, Inc.

Artificial intelligence (AI) is reshaping how environmental professionals manage, interpret, and act on complex datasets. The EQuIS platform integrates AI-powered tools—most notably EQuIS Helios—with strong data governance, automated workflows, and robust knowledge management capabilities. This combination improves remediation strategies, enhances decision-making, and supports sustainable environmental outcomes.

EQuIS Helios uses Microsoft AI services to extract insights from unstructured data sources such as PDFs, field notes, images, and scanned documents. It enhances searchability, summarizes content, and flags sensitive information such as personally identifiable information (PII) and profanity.

Helios supports a range of real-world use cases, to name a few:

• Government and industrial users consolidate decades of consultant reports and field notes. Helios indexes, summarizes, and secures this information for centralized access.
• Consultants process legacy project data from spreadsheets, PDFs, and scans. Helios summarizes and prepares these datasets for review and potential migration into EQuIS.
• Organizations manage historical archives such as boring logs and permits. Helios makes these records searchable by location, metadata, and content.
• Field teams extract data from handwritten notes and scanned forms. Helios structures this data for validation and integration into broader workflows.

The EQuIS platform enforces data quality through enterprise standards and automated validation. These controls ensure high-quality, reliable data that supports accurate predictions, risk identification, and pollution pattern detection. By embedding AI into a federated data environment, EQuIS enables users to access and analyze data across distributed systems without requiring duplication. This approach improves data visibility, accelerates response times, and supports smarter, more adaptive remediation strategies.

Coauthor: Dan Alexander (EarthSoft, Inc.)

4:15 - 4:35 p.m.

Leveraging Artificial Intelligence and Large Language Models for Enhanced Decision-Making in Environmental Remediation

Inci Demirkanli, Pacific Northwest National Laboratory

Environmental remediation (ER) of radioactively contaminated sites can be a complex endeavor, with significant uncertainty and many factors to consider, involving iterative evaluation of a multitude of technical, environmental, economic, and community factors, as well as regulatory considerations and requirements, to determine a remedial approach that mitigates contaminant risk and maximizes the benefits and acceptability.  As defined by the International Atomic Energy Agency’s (IAEA), remediation includes any measures that may be carried out to reduce radiation exposure from existing contamination of land areas through actions applied to the contamination itself (the source) or to exposure pathways to humans. It is also important that design and implementation of site remedial management activities optimize the net benefits of economic, environmental, and social gains for the present and into the future.

Data-driven approaches, especially those utilizing artificial intelligence and machine learning (AI/ML), are invaluable for navigating this complexity. By leveraging the vast accumulated knowledge from historical records, diverse databases, regulatory frameworks, and lessons learned, AI/ML can support the identification of potential remedial alternatives and lead to technically defensible and broadly acceptable decision outcomes.

Building on these principles, we explore the specific application of Large Language Models (LLMs)—a powerful advancement in AI—to enhance the ER decision-making process. Drawing from a blueprint decision support framework successfully demonstrated in the environmental permitting domain (PermitAI), we examine how LLM technology can augment human workflows in remediation. Specifically, this technology can synthesize information from heterogeneous sources, help identify appropriate site assessment strategies, evaluate safety and environmental impacts, and ultimately aid in the planning and selection of remedial options. By demonstrating a tangible framework that facilitates dialogue and promotes interdisciplinary collaboration, this approach aims to educate the community and support stakeholders, including IAEA Member States, in achieving more efficient and effective environmental cleanup.

Coauthors: Sai Koneru, Chris Johnson, Sai Munikoti, and Sameera Horawalavithan (PNNL); Horst Monken Fernandes (IAEA)