Predicting swelling pressure of bentonite and bentonite mixtures using various machine learning approaches

Abstract

Bentonite and bentonite mixtures are used as buffer material for deep geological radioactive waste repositories. The swelling behavior of bentonite is an important property influencing the long-term safety of the barrier system by its self-sealing effect. The proper determination of bentonite swelling pressure is vital to ensure that geological repositories remain intact. In this study, a total of 305 data samples on bentonite swelling pressure was collected from the literature. Corresponding soil properties were montmorillonite content, liquid limit, plastic limit, plasticity index, initial water content, and dry density. We employed various machine learning algorithms, namely feed-forward and cascade forward neural networks, regression tree, regression tree ensembles, Gaussian process regression, and support vector machines to determine the maximum swelling pressure of unsaturated bentonite and its mixtures. The cascade-forward neural network (CFNN) produced the best overall performance, i.e. the lowest modeling deviations from the experimental swelling pressure values. Furthermore, we present two simplified CFNN models that depend on two (montmorillonite content and initial dry density) and three (montmorillonite content, initial dry density, and plasticity index) variables to estimate bentonite swelling pressures. These simplified models can to be used as alternatives in instances of limited data availability.