Machine learning and regression-based approaches for UCS estimation in weathered sandstone

The uniaxial compressive strength (UCS) of weathered sandstone was estimated in this study using machine learning (ML) and regression analysis, integrating pulse velocity (UPV), bulk density, and weathering grades of sandstone samples. The weathered sandstone specimens are classified from slightly (II) to highly (IVb) weathered and tested under unconfined compressive test and portable ultrasonic non-destructive indicating tester (PUNDIT) to determine the UCS and UPV values. A total of 79 samples spanning Grades II–IVb were tested in accordance with the international society for rock mechanics (ISRM). The experimental results indicate that UCS, UPV, and bulk density values of sandstone decreased by 69.7%, 57.8%, and 15.8% for IVb grade compared to II grade. The complex, nonlinear, and interpretable relations were captured using the random forest method (with out-of-bag validation) and grade-wise ordinary least-squares regressions. The results imply that UPV is a dominant predictor of UCS, with density providing complementary information; weathering grade retained moderate importance, reflecting macro-scale degradation not captured by intrinsic properties alone. The models achieved high predictive fidelity for slightly to moderately weathered samples, with reduced accuracy in highly weathered rock where micro-crack connectivity and textural anisotropy prevail. The hybrid approach yielded high predictive accuracy for slightly to moderately weathered samples (pearson correlation coefficient (r) ≥ 0.89), with reduced performance for highly weathered sandstone due to increased heterogeneity. The proposed workflow enables rapid, reliable, and field-deployable strength estimation, supports grade-aware design values, and reduces reliance on destructive testing, thereby advancing data-driven geotechnical classification and preliminary design in tropical weathered formations.