Artificial intelligent modelling of uniaxial compression strength of rock using machine learning algorithms
摘要
Accurate prediction of uniaxial compressive strength (UCS) is essential for rock engineering design, while existing empirical formulas often exhibit limited accuracy and generalization capability. This study proposes a machine learning–based surrogate modeling framework to predict UCS using different combinations of commonly measured rock properties. A comprehensive dataset collected from published literature, including porosity (n), Schmidt hammer rebound number (SH), P-wave velocity (Vp), and point load index (Is), was used to develop UCS prediction models. Eleven feature combinations were constructed and implemented using the XGBoost algorithm. The dataset was divided into training and testing subsets, and model performance was evaluated through cross-validation and independent testing. Multiple indicators, including RMSE, MAE, MAPE, and Spearman correlation coefficients, were employed to assess prediction accuracy and robustness. The results indicate that models incorporating porosity consistently outperform those without it. Among all feature combinations, n-SH, n-SH-Vp, and n-SH-Vp-Is models show the best performance in the two-, three-, and four-variable groups, respectively, and are recommended for practical UCS prediction depending on data availability. Parametric investigation reveals smooth and physically meaningful relationships between UCS and input variables. Sobol-based sensitivity analysis demonstrates that porosity has the dominant influence on UCS, whereas the contribution of the point load index is relatively minor. Comparisons with commonly used empirical formulas and uncertainty evaluation using prediction intervals further confirm the reliability and superior predictive capability of the proposed surrogate models for rock engineering applications.