Estimation of Intact Rock Uniaxial Compressive Strength Using Advanced Machine Learning

The present investigation introduces an optimal computational model by comparing gene expression programming (GEP), least square support vector machine (LSSVM), and extreme learning machine (ELM) models in predicting the uniaxial compressive strength of intact rocks (RUCS). This research employs linear, polynomial, and radial basis function (RBF) kernel-based LSSVM models and compares them. Furthermore, this investigation reveals the effect of chromosomes on the performance of GEP models. One hundred four and 27 results of RUCS have trained and tested RUCS models. In addition, the multicollinearity has been computed for an overall database to analyze its effect on the model’s performance and accuracy. This research uses the area and mass of rock specimens, along with Young’s modulus, for the first time in predicting the rock UCS. The performance metric comparison reveals that model ELM (mentioned by RUCS7) has predicted rock UCS with a correlation coefficient of 0.9642, root mean square error of 0.0479 MPa, performance index of 1.8067, variance accounted for of 92.49, and agreement index of 0.8681, comparatively higher than LSSVM and GEP models and close to ideal values. The performance analysis reveals that weak multicollinearity affects the prediction capabilities of the linear LSSVM model. Conversely, it has been observed that the multicollinearity effect can be controlled at a certain level by implementing more chromosomes in the GEP model. The area, mass, and Young’s modulus of rock specimens highly influence the prediction of RUCS. Finally, the Wilcoxon test (confidence interval = 0.1450), uncertainty analysis (rank = 1), score analysis (overall = 238), and generalizability (≈ ideal values) demonstrate the ELM model as an optimal computational model for assessing the rock UCS.