A BOHB-optimized stacking ensemble approach for predicting the coefficient of compressibility and quantifying uncertainty of guilin lateritic clays

The coefficient of compressibility of lateritic clays is central to settlement analysis and foundation design, yet rapid assessment is hampered by their paradoxical combination of high plasticity and low-to-medium compressibility. A Bayesian Optimization with HyperBand (BOHB)-optimized stacking ensemble (CatBoost-XGBoost-LightGBM + ElasticNet) predicts this coefficient from basic physical indices using 454 datasets from Guilin, China. The model achieves a test R 2 of 0.9692 and RMSE of 0.0183 MPa −1 , outperforming all individual base models. SHAP and partial dependence analyses reveal that liquidity index and water content dominate the prediction, accounting for over 95% of feature importance, with their synergistic elevation as the primary driver of high compressibility. Spatial cross-validation yields a mean test R 2 of 0.9415, and Monte Carlo simulation confirms robustness: 80% of samples yield a 90% confidence interval width below 0.05 MPa −1 . This framework enables rapid, reliable compressibility assessment for lateritic clays.