AN EXPLAINABLE MACHINE LEARNING FRAMEWORK FOR PREDICTING UNIAXIAL COMPRESSIVE STRENGTH OF ROCK USING INDEX PROPERTIES

Abstract

The uniaxial compressive strength (UCS) of rock is one of the most important engineering properties used in the design and analysis of tunnels, slopes, foundations, underground excavations, and mining projects. Although the laboratory determination of UCS provides reliable results, it is often time-consuming, expensive, and requires carefully prepared standard specimens. In many practical situations, obtaining suitable rock samples for direct UCS testing is difficult. Consequently, the prediction of UCS using easily measurable index properties has become an important area of research in rock mechanics and geotechnical engineering. This study presents an explainable machine learning framework for predicting the UCS of rocks using four commonly measured index properties: Porosity (%), Schmidt Hammer Rebound (SHR), P-wave velocity (Vp), and Point Load Index (Is50). A dataset of 584 rock specimens, compiled primarily from the open-access ROCK/10/4025 database and supplemented with specimen-level data reported by Abd El Aal (2017), was used for model development following a zero-imputation policy that retained only specimens with complete measurements. Three machine learning algorithms which are Random Forest Regression, Ridge Regression, and Huber Regression were developed and evaluated on an 80:20 train-test split using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). The Random Forest Regression model achieved the best performance on the test partition, with an R² of 0.73, an RMSE of 26.19 MPa, and an MAE of 16.96 MPa, outperforming both linear alternatives. To improve the transparency and interpretability of the predictive model, SHapley Additive exPlanations (SHAP) analysis was applied to the trained Random Forest model to quantify the contribution of each index property to the predicted UCS. The explainability analysis, consistent with the Pearson correlation trends observed in the dataset, ranked porosity as the most influential parameter, followed by P-wave velocity, Schmidt Hammer Rebound Number, and Point Load Index. Based on the trained models, an empirical equation for UCS estimation was also developed to provide a practical, rapid engineering tool. iv The developed framework was further validated against ten independently tested laboratory rock specimens that were not part of the training dataset. Since the Point Load Test and the UCS Test are both destructive and cannot be performed on the same specimen, the corresponding Is50 values for the validation samples were estimated theoretically using established empirical correlations. The comparison between laboratory-measured and model-predicted UCS values showed close agreement, with an MAE of 3.09 MPa, an RMSE of 4.00 MPa, and a mean absolute percentage error (MAPE) of 10.32% across the ten specimens. Prediction accuracy was found to be highest for competent, low-porosity rock and comparatively lower for weaker, more porous specimens, where micro-scale heterogeneity has a greater relative influence on strength. As this validation exercise was based on a limited set of ten specimens, it should be regarded as an encouraging preliminary indication rather than conclusive proof of the framework's reliability, and broader validation on larger, independent datasets is recommended before wider adoption. The proposed explainable machine learning framework offers an efficient, accurate, and interpretable approach for estimating the uniaxial compressive strength of rocks. It has the potential to reduce the need for extensive laboratory testing and can assist geotechnical engineers in preliminary site investigations, rock mass characterisation, and engineering design where rapid and reliable UCS estimation is required.