Abstract: High-level radioactive waste repository has a long service period and complex geological conditions. It is difficult for general tests to obtain the evolution trend of the swelling performance of bentonite buffer barrier under the working conditions of the repository. This paper attempts to use machine learning algorithms to research it. The model is trained by ten machine learning algorithms which is more mature in material performance prediction, the coefficient of determination and root mean square error of the above algorithms are compared and analyzed. The feasibility of machine learning algorithms in predicting the swelling performance of bentonite is discussed, and the neural network algorithm is selected as the best one. The geochemical condition of Beishan candidate repository in China is introduced. It is predicted that the swelling pressure evolution trend of Gaomiaozi bentonite during the service of high-level radioactive repository is ‘S’ type. The results show that considering the influence of dry density, montmorillonite content, initial moisture content and temperature, the recommended value of initial dry density of buffer barrier is not less than 1.58 g/cm³. If the effect of Fe(II) release from the iron container is further considered, the initial dry density of the buffer barrier should be further increased. This study can provide a theoretical basis for the in-situ experiment and buffer barrier performance evaluation of the underground laboratory of Beishan high-level waste radioactive repository in China.

Key words: bentonite, high-level radioactive waste repository, buffer barrier, swelling performance, machine learning