Predicting peak shear strength of rock fractures using tree-based models and convolutional neural network

It is of significance to estimate peak shear strength (PSS) of rock fractures in engineering practice, but the existing PSS criteria may not fully represent the 3D characteristics of fracture surfaces. In this study, data-driven PSS criteria of rock fractures are developed to comprehensively consider effects of surface roughness features. An effective method to create a large high-quality dataset is first proposed by combining particle-based discrete element method for numerical shear tests with diamond-square algorithm for generating random fracture sur-faces. Five tree-based models are trained based on the dataset containing normal stress, rock mechanical properties and 16 explicit roughness features, while convolutional neural network is trained based on the mixed dataset containing normal stress, rock mechanical properties and relative fracture elevation. The prediction accuracy of the trained data-driven PSS criteria is examined using additional experimental data, and the results show that the tree-based models of categorical boosting and light gradient boosting machine and convolutional neural network can provide reliable prediction of PSS of rock fractures. The dominant features are ranked ac-cording to their contributions to the tree-based PSS criteria. The data-driven PSS criteria of rock fractures would have a great potential in engineering application with limited access to experimental data.

Automated summary

This study develops data-driven criteria to estimate the peak shear strength (PSS) of rock fractures, considering the effects of surface roughness features. A high-quality dataset is created using particle-based discrete element method and diamond-square algorithm. Tree-based models and convolutional neural network are trained to predict the PSS of rock fractures, and their reliability is verified using experimental data.