Applying machine learning methods to estimate the thermal conductivity of bentonite for a high-level radioactive waste repository

Abstract

Bentonite is an important engineering barrier in the deep geological repository of high-level radioactive waste (HLW), and its thermal conductivity is one of the key parameters in the design of the multi barrier system for HLW repository. In this study, several machine learning (ML) algorithms and two input set (3-variable and 5-variable) were used to estimate the thermal conductivity and thermal diffusivity of bentonite. After training and comparing, the models obtained from Random Forest, Extra Trees Regressor and Gradient Boosting algorithms show better performance. The generalizability of these models was verified by the thermal conductivities of Gaomiaozi bentonite and Gyeongju bentonite, and the result shows that the ML models with 5 input variables have better generalizability. The influence and importance of input variables were also discussed. ML models could provide clues to improve the thermal conductivity of bentonite in engineering, and have great potential to evaluate the long-term behavior of bentonite under the condition of HLW repository.