The distribution coefficient (K_d) of radionuclides on bentonite is one of the key parameters in the safety assessment of geological repositories for high-level radioactive waste.

This study aims to reliably predict the K_d values of various radioactive cations in bentonite based on machine learning (ML) models.

Based on the Japan Atomic Energy Agency's adsorption database (JAEA-SDB), 1 240 sets of K_d data of 10 nuclides (Am, Bi, Cm, Cs, Eu, Ni, Pb, Po, Ra, U) in bentonite were collected, and 9 input factors were selected to construct six ML models including random forest (RF) and support vector regression (SVR). The robustness of the ML models was evaluated by Monte Carlo cross-validation (MCCV).

The validation results indicate that the RF model is the most effective in predicting K_d, achieving a determination coefficient (R²) of 0.902 9 on the training set and 0.728 6 on the test set. Moreover, the RF model accurately reproduce the probability density distribution characteristics of K_d, and pH, initial concentration, ionic strength, and temperature are the main factors affecting K_d.

Results of this study demonstrate that ML methods, especially the RF model, can rapidly and reliably predict the K_d of multiple radionuclides on bentonite under complex conditions, offering a promising new approach for the safety assessment of radioactive waste disposal repositories.