| Abstrakt: |
This study aims to make a unique contribution to the existing body of knowledge about rock strength and deformation parameters and crack stress thresholds through intelligent and statistical approaches applied to a database comprising various rock types (i.e., sedimentary, igneous, and metamorphic rocks). The database contains physical–mechanical and ultrasonic parameters. Six distinct machine learning (ML) algorithms— artificial neural network (ANN), random forest (RF), decision tree (DT), K-nearest neighbor (KNN), support vector regression (SVR), and bagging regressor (BR)— along with the conventional linear regression techniques, were employed to develop predictive models. These models estimate uniaxial compressive strength ( σ c ) and Tangent Young's modulus ( E t ) based on bulk density (ρ ) and P-wave ultrasonic velocity ( V p ). Furthermore, they predict crack stress thresholds (i.e., crack closure stress σ cc , crack initiation stress σ ci , and crack damage stress σ cd ) as a function of σ c , E t , ρ , V p , axial strain at failure ( ε 1 f ), and lateral strain at failure ( ε 3 f ). Various performance indices were utilized to evaluate and compare the performance of these models. The results indicated that the RF method outperformed other ML-based and linear regression-based approaches in predicting the output parameters. Additionally, the multiple parametric sensitivity analysis (MPSA) was carried out to determine the significance of input parameters in predicting the output variables. This analysis revealed that V p and ρ have the highest and lowest impact on predicting σ c and E t , respectively. On the other hand, σ c was identified as the most influential parameter in predicting σ ci and σ cd , while parameters ε 3 f and V p showed the least impact on the foregoing outputs, respectively. This is while ε 1 f and ρ were, respectively, found as the most important and least important factors in predicting σ cc . Finally, to facilitate easy access to the prediction results and enhance the practicality of the proposed RF model, a graphical user interface (GUI) was developed, which enables the practical application of the most performing developed prediction model. Highlights: A comprehensive database including different rock types was prepared to evaluate the rock strength parameters and crack stress thresholds. The used six robust machine learning algorithms effectively unveiled the latent complex nonlinear relationship between the parameters. The random forest algorithm outperformed the traditional regression and machine learning algorithms in predicting the output variables. P-wave velocity, uniaxial compressive strength, and axial strain at the failure point were identified as the most influential parameters in predicting outputs. The developed user-friendly graphical user interface brings the power of the random forest model to practitioners without the need for complex analyses and destructive laboratory tests. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink