Calibration of the Microparameters of Rock Specimens by Using Various Machine Learning Algorithms

High accuracy in the simulation of the discrete-element method (DEM) depends on the proper selection of microparameters. In this study, the range of microparameters was determined through sensitivity analysis. Subsequently, four levels of orthogonal experimental tables were established and 148 sets of data were collected. In addition, five data mining methods, namely, support vector regression (SVR), nearest-neighbor regression (NNR), Bayesian ridge regression (BRR), random forest regression (RFR), and gradient tree boosting regression (GTBR), were used to establish a microparameter prediction model. The results indicate that machine learning methods have significant potential in determining the relationship between macro and microparameters of the DEM model. RFR achieved the best performance among the five models whether the input data were collected from the tests of the Brazilian tensile strength and uniaxial compression or only the uniaxial compression test. In addition, the deviation between the predicted and measured macroparameters was less than 8%. This approach allowed for more accurate modeling of complex structures in a rock under various stress conditions through DEM simulations.

Automated summary

This study determined the range of microparameters in the discrete-element method through sensitivity analysis and established a microparameter prediction model using five data mining methods. The results suggest that machine learning methods have significant potential in determining the relationship between macro and microparameters of the DEM model, with RFR model achieving the best performance and the deviation between predicted and measured macroparameters being less than 8%.