A constraint-enhanced machine learning model for predicting hydraulic conductivity of unsaturated bentonite

The accurate determination of hydraulic conductivity in unsaturated bentonite is important for the modeling of subsurface thermo-hydro-mechanical and chemical processes. This study introduces a new hybrid approach, employing a constrained CatBoost algorithm coupled with a genetic algorithm for hyperparameter tuning. We benchmarked the effectiveness of the constrained CatBoost model against various data-driven regression models, including lasso, elastic net, polynomial regression, k-nearest neighbors, decision tree, bagging tree, random forest, and standard CatBoost. Our findings demonstrate that the constrained CatBoost model excels in providing accurate estimations of the hydraulic conductivity of compacted bentonite during the wetting phase. The model adequately captures the U-shaped correlation between hydraulic conductivity and suction and reflects the influence of temperature changes on hydraulic conductivity. Furthermore, the bootstrapping analysis, conducted across 800 iterations, confirms the stability and robustness of the constrained CatBoost model. This work provides a reliable tool for predicting hydraulic conductivity in diverse environmental and engineering contexts.