The determination of fundamental rock mechanical properties, uniaxial compression strength (UCS) and elastic modulus (E), constitutes a pivotal facet of rock engineering design. However, deriving these properties directly from standard laboratory tests on rock core samples can be challenging, especially when dealing with deep high-stress rock formations and weak fractured strata. Thus, it is crucial to establish a cost-effective and practical approach for predicting the macro-mechanical properties of rocks in situ. In this study, a machine learning approach was proposed to predict UCS and E by upscaling meso-mechanical parameters at particle scale in low-porosity crystalline rocks. To expand the correlation database of rock meso-macro mechanical properties, the meso-mechanical parameters, including the fracture toughness, tensile strength of the rock crystal interface, and the elastic modulus of rock crystal, were accurately measured, using a newly designed mechanical apparatus and a nanoindentation device. The grain-based models implemented in the combined finite discrete element method (FDEM-GBM) were developed based on these experimental results, and their reliability was validated though standard tests. Subsequently, a database, including 225 groups of data, was established using the numerical method. Five machine learning algorithms were applied to develop prediction models for UCS and E through data training in the database. Excellent performance improvement was achieved through the application of the grid-search method. The results indicate the optimized kernel ridge regression (KRR) and gaussian process regression (GPR) models demonstrated excellent performance with relative average errors of 4.9% and 1.1% in predicting UCS and E, respectively. Finally, the predicted values of UCS and E were compared with the experimental results, validating the feasibility of the optimized KRR and GPR models.
Read full abstract