An explainable ensemble machine learning approach for coefficient of compressibility prediction of Chengdu clay

Accurate prediction of the coefficient of compressibility (a v ) for regional clays remains a fundamental challenge in geotechnical engineering. Focusing on Chengdu clays, this study proposes a machine learning framework that integrates physically constrained data augmentation, differentiated hyperparameter optimization, and interpretability analysis. Based on 338 experimental datasets, a hybrid optimization strategy (Grid Search + Bayesian Optimization) is employed to ensure fair model comparison. The statistical significance of performance differences is rigorously validated using 30 iterations of Monte Carlo cross-validation combined with paired t-tests. Generalization capability is assessed via spatial grouping cross-validation, accompanied by Bootstrap-based uncertainty quantification. Results demonstrate that LightGBM achieves the highest predictive accuracy (test R 2 = 0.8817), significantly outperforming comparison models (p < 0.05). XGBoost exhibits comparable accuracy (R 2 = 0.8805), while Random Forest demonstrates superior robustness in spatial cross-validation (R 2 = 0.7273). Multi-model analysis consistently identifies water content and liquidity index as the dominant factors governing a v , with cumulative contribution exceeding 95%. SHAP analysis confirms both factors exhibit monotonically positive effects with approximate additivity. Tree-based models yield Bootstrap confidence intervals with mean widths ranging from 0.0526 to 0.0653 MPa -1 , providing essential risk references for engineering practice. This study not only delivers a reliable predictive tool for a v of Chengdu clays but also establishes a reproducible, physically guided, and statistically rigorous paradigm for regional geotechnical machine learning research.