Prediction and interpretability analysis of hydraulic conductivity for polymer-amended soil-bentonite cutoff wall backfills using machine learning methods

Soil-bentonite backfills with low hydraulic conductivity are extensively used to control flow of groundwater at contaminated sites. Predicting hydraulic conductivity of backfills is challenging due to nonlinear couplings among chemistry reactions, backfill compositions, and in-situ stresses. This study developed a machine-learning framework that unifieed a curated dataset of 108 laboratory observations, systematic benchmarking of twelve algorithms, SHAP-based interpretability, and mechanism-guided feature subset optimization to balance accuracy with measurement cost. Validation used an 80/20 training/testing split with five-fold cross-validation for hyperparameter tuning, multi-metric evaluation, and a Taylor diagram for consistency appraisal. Peak test performance showed clear stratification: Gradient Boosting Regressor (R² = 0.794), Support Vector Regression (R² = 0.788), XGBoost (R² = 0.783), Random Forest (R² = 0.736), LightGBM (R² = 0.698). Taylor diagram corroborated this ordering by proximity to the reference point. SHAP analysis identifieed leachate pH as the top contributor in Gradient Boosting Regressor, and explained why low ionic strength lowered the predicted hydraulic conductivity whereas higher ionic strength increased it. Feature-count analysis indicateed that a stable baseline requireed at least 5 features and that gains generally tapered after 9 to 10 features. The results motivated a three-tier workflow: rapid screening with a compact interaction set, routine design with curated features, and final validation with the all features set.