Predicting distribution coefficient and effective diffusion coefficient of radionuclides in bentonite: Multi-output...

Area of Science:

• Geochemistry and Environmental Science
• Artificial Intelligence and Machine Learning
• Nuclear Engineering and Safety

Background:

• Accurate prediction of radionuclide transport is crucial for the safety assessment of high-level radioactive waste (HLW) repositories.
• Distribution coefficient (K_d) and effective diffusion coefficient (D_e) are key parameters governing radionuclide migration.
• Existing models often require extensive experimental data and may not capture complex interactions effectively.

Purpose of the Study:

• To develop a novel, integrated Artificial Neural Network (ANN) framework for simultaneously predicting both K_d and D_e.
• To enhance the predictive capability by augmenting the dataset with synthetic data generated via Generative Adversarial Network (GAN).
• To identify key geological factors influencing radionuclide transport through explainable AI techniques.

Main Methods:

• A multi-output ANN model was designed for simultaneous prediction of K_d and D_e.
• Generative Adversarial Network (GAN) was utilized for data augmentation, creating pseudo-instances to expand the training dataset.
• Shapley Additive Explanations (SHAP) analysis was employed to interpret model predictions and identify influential features.

Main Results:

• The GAN-ANN model achieved high predictive accuracy, with R² values of 0.98 for K_d and 0.97 for D_e.
• Total porosity was identified as the most significant predictor for both K_d and D_e.
• Experimental validation using through-diffusion tests on various radionuclides in compacted bentonite and I/S confirmed the model's robust generalization capability.

Conclusions:

• The developed GAN-ANN model provides a powerful and accurate tool for predicting radionuclide sorption and diffusion in HLW repositories.
• The study demonstrates the effectiveness of combining GAN for data augmentation with ANNs for multi-output prediction in complex geoscience applications.
• The findings contribute to a more reliable dataset and predictive framework, significantly aiding in the safety assessment of nuclear waste disposal.