Application of machine learning methods to predict a thermal conductivity model for compacted bentonite

Abstract

Compacted bentonite is one of the widely used buffer materials for the disposal of high-level radioactive waste (HLW). Since the buffer is located between a disposal canister and near-field rock, it prevents the release of radionuclides, and protects the canister from external impact and the penetration of groundwater. To dissipate decay heat from the canister, the buffer should have high thermal conductivity, so one of the most important properties for HLW repository is the thermal conductivity of compacted bentonite. In this study, predictive models of thermal conductivity for the compacted bentonite has been designed and analyzed using machine learning methods including linear regression, decision tree, support vector machine, ensemble, Gaussian process regression (GPR), artificial neural network, and deep belief network. Most of the methods showed better performance in comparison with the previously proposed regression model, while the GPR with exponential kernel and the ensemble with XGBoost showed the best performance.