A physics-based CatBoost model for water retention of compacted bentonite with global sensitivity analysis

Bentonite is a recommended buffer material in high-level radioactive waste repositories to restrict the migration of radionuclides into the environment. Determining the soil water retention curve (SWRC) of bentonite is essential for predicting its hydraulic behaviour, including water flow dynamics and saturation time, which are critical for evaluating the performance of engineered barrier systems. This study compiled 46 experimental SWRCs from existing literature containing 311 data points of matric potential and corresponding water content. Key soil properties associated with these data points include specific gravity, montmorillonite content, initial dry density, initial water content, initial void ratio, and plasticity index. The Van Genuchten model parameters were optimized using the Levenberg–Marquardt algorithm for each of the 46 SWRCs. To enrich the SWRC data, 20 additional data points of matric potential were generated, and the predicted water content from the optimized Van Genuchten models was then combined with the experimental data. A machine learning model was developed to predict the SWRC of bentonite using the CatBoost machine learning algorithm; and fine-tuned its hyper-parameters using the artificial gorilla troops optimizer. As input, the machine learning model used matric potential, key soil properties, and experimental conditions such as confined or unconfined states and drying or wetting paths. The machine learning model shows very good performance in estimating the water content at various matric potentials, offering an efficient method to determine the SWRC of bentonite based on key soil properties. • Developed a machine learning model to predict the SWRC of bentonite accurately. • Used data augmentation with the Van Genuchten model to enhance predictive accuracy. • Sensitivity analysis shows matric potential, confinement, and montmorillonite are key factors. • Introduced streamlined machine learning models for limited soil data scenarios.