Developing a machine learning-based predictive model for cesium sorption distribution coefficient on crushed granite.

Area of Science:

• Geochemistry
• Environmental Science
• Materials Science

Background:

• Granite's radionuclide sorption is vital for high-level radioactive waste (HLW) geological disposal safety assessments.
• Sorption properties of granite show significant variability under diverse experimental conditions.
• Traditional experimental methods for assessing sorption are time-consuming and costly.

Purpose of the Study:

• To develop an efficient, data-driven machine learning (ML) approach for predicting cesium (Cs) sorption distribution coefficients on crushed granite.
• To reduce the time and cost associated with traditional experimental sorption studies.
• To identify key factors influencing Cs sorption on granite.

Main Methods:

• Utilized four ML algorithms: AdaBoost, GBDT, LightGBM, and XGBoost.
• Trained models on a dataset of 384 experimental data points.
• Employed SHAP analysis to interpret model predictions and feature importance.

Main Results:

• All ML models achieved R-squared values greater than 0.8 for both training and testing datasets.
• XGBoost demonstrated superior predictive performance and generalization ability.
• Key features influencing Cs sorption were ranked: solid/liquid ratio, ion strength, pH, contact time, initial concentration, and maximum particle size.

Conclusions:

• The developed ML approach effectively predicts radionuclide sorption on granite, complementing resource-intensive experiments.
• SHAP analysis revealed sorption mechanisms consistent with experimental observations.
• This data-driven method enhances the safety assessment of HLW geological disposal by providing insights into radionuclide behavior.