Explainable ensemble learning for estimating strength and stiffness of clay-bearing rocks: A case study from the...

Reliable estimation of unconfined compressive strength (UCS) and Young's modulus (E) is critical for geotechnical design in weak, clay-rich formations. This study develops an integrated experimental-machine learning (ML) framework to predict UCS and E for marl and shale from the Godarkhosh Dam site (Iran). Four ensemble ML algorithms (LightGBM, AdaBoost, CatBoost, Random Forest) are trained using non-destructive indices (density, porosity, water absorption, and P-/S-wave velocities). AdaBoost achieves the highest UCS prediction performance (test R² = 0.90, RMSE = 0.70 MPa), while CatBoost provides the most accurate E estimates (test R² = 0.82, RMSE = 0.22 GPa). Explainable AI (SHAP, LIME, partial dependence plots) consistently identifies seismic velocities and dry density as dominant positive predictors, and porosity and water absorption as negative influences. Monotonic, physically coherent trends across all XAI tools confirm that the models capture meaningful geomechanical relationships rather than statistical artefacts. The proposed ML-XAI workflow enables accurate, interpretable estimation of mechanical properties in soft rock terrains using a minimal set of rapid, non-destructive measurements, supporting efficient site characterization for dam foundations and similar weak sedimentary formations.