Physics-based and data-driven machine learning modeling of saturated and unsaturated hydraulic conductivity of bentonite

Accurate prediction of the hydraulic conductivity of compacted bentonite is critical for assessing the long-term safety of high-level radioactive waste repositories, where barrier efficiency depends on coupled processes. This study develops a data-driven machine learning model to predict saturated hydraulic conductivity and a physics-based machine learning model to predict unsaturated hydraulic conductivity of compacted bentonite. For the saturated hydraulic conductivity prediction, a dataset of 215 experimental measurements was compiled, incorporating key soil properties such as montmorillonite content, specific gravity, plasticity index, initial water content, dry density, and temperature as input. To predict unsaturated hydraulic conductivity, the study considers experimental data, synthetic data generated using the Van Genuchten model, and outputs from the machine learning model developed for saturated hydraulic conductivity. The input dataset includes specific gravity, montmorillonite content, initial dry density, initial water content, initial void ratio, plasticity index, and suction. The AdaBoost, CatBoost, and XGBoost algorithms were used to train the machine learning models, and the whale optimization algorithm was used for hyperparameter tuning. The trained machine learning models demonstrate good predictive performance for both saturated and unsaturated hydraulic conductivity of compacted bentonite, showing close agreement with experimental measurements.