Comparison between some methods for predicting the fracture of hard rock in three-point bending tests. (February 2023)

• Record Type: Journal Article
• Authors: Han, Xiangyu; Xiao, Qinghua; Cui, Kai; Lei, Shengxiang; Hu, Xiaozhi; Chen, Shougen
• Is Part Of: Theoretical and applied fracture mechanics. Volume 123(2023)
• Publication Date: 2023-02
• Subjects: Hard rock -- Fracture behavior -- Three-point bending test -- Prediction approaches; Fracture mechanics -- Periodicals; 620.1126
• DOI: 10.1016/j.tafmec.2022.103689
• Languages: English
• ISSNs: 0167-8442

Abstract

Highlights: Fracture prediction of hard rock specimens with three-point bending tests is investigated. Specimen geometries have impact on accuracy of linear elastic fracture prediction model. The non-linear elastic fracture model performed better in fracture prediction. The artificial intelligence method is suggested to be assisted with mechanical theory. Abstract: Fracture properties are fundamental for hard rock materials but cannot be precisely described by conventional strength theories, which makes the prediction of fracture behavior of hard rock troublesome. With the development of fracture mechanics and numerous interdisciplines, various novel methods are put forward to solve the fracture problems of hard rocks. In this study, three-point bending fracture test results of two types of hard rock are presented, two mechanical methods and one artificial intelligence approach are attempted to predict the fracture performance of rock specimens. Randomly selecting parts of test data to train the predictive models, and using the rest data to test the predictive performance of models. From the comparison between these approaches, the Boundary Effect Fracture Model (BEM) turns out to be a better mechanical method for predicting the fracture of hard rock, the size effect of fracture properties and the fracture prediction of specimens without initial notch are also well treated by the Boundary Effect Fracture Model, the predictive performance is still good with limited training data. The artificial intelligence (AI) method also has excellent predictive behavior ( R 2 = 0.979) by combining the gradient boosting regression tree and firefly algorithms, but the obvious prediction error will appear when some training data is missing.