Predicting radionuclide behavior in deep geological repositories using graph convolutional long short-term memory.

Area of Science:

• Geological engineering
• Computational modeling
• Nuclear waste management

Background:

• Deep geological repositories (DGRs) require accurate radionuclide transport predictions for safe spent nuclear fuel disposal.
• Physical experiments are impractical for long-term simulations, making computational models essential.
• The Parallel Flow and Reactive Transport Model (PFLOTRAN) is a standard but computationally intensive tool.

Purpose of the Study:

• To develop and validate a Graph Convolutional Long Short-Term Memory (GCLSTM) model as a computationally efficient surrogate for PFLOTRAN.
• To assess the accuracy and reliability of GCLSTM for simulating long-term radionuclide transport in DGRs.
• To reduce the computational burden of radionuclide transport simulations.

Main Methods:

• GCLSTM was trained using time-series data from 5,000-year PFLOTRAN simulations.
• Model performance was evaluated using coefficient of determination and Nash-Sutcliffe efficiency.
• Uncertainty quantification, sensitivity analyses, and scenario-based simulations were performed.

Main Results:

• GCLSTM achieved a coefficient of determination >0.99 and Nash-Sutcliffe efficiency >0.97.
• Over 95% of GCLSTM predictions fell within PFLOTRAN confidence intervals.
• Permeability and inter-node distance were identified as key drivers of predictive variance.

Conclusions:

• GCLSTM serves as an efficient and accurate surrogate for PFLOTRAN in radionuclide transport simulations.
• The GCLSTM model reduces computational time by approximately 576 times.
• This approach offers a practical solution for enhancing modeling efficiency in DGR safety assessments.