Machine learning-aided prediction of the mechanical properties of frozen fractured rocks

Abstract

The complexity of fracture geometries impedes reliable prediction of the mechanical properties of frozen fractured rocks. Here, we combine the experimental, numerical, and machine learning methods to predict the uniaxial compressive strength and the Young’s modulus of frozen fractured rocks with five fracture geometries, including persistence factor of ice-filled fractures, spacing between the fractures, as well as inclination angle, thickness, and number of the fractures. We use the results of laboratory uniaxial compression tests to validate the numerical model and the results of two-dimensional particle flow code simulations to train the random forest (RF) models. Our study demonstrates reliable prediction of the uniaxial compressive strength and the Young’s modulus of frozen fractured rocks and compares the prediction performance with the Ramamurthy criterion. We also conduct a sensitivity analysis to reveal dominant geometries and obtain the simplified RF models with three fracture geometries (i.e., persistence factor, inclination angle, and fracture number) for similar prediction accuracy. We finally use additional experimental results to further test the reliability of the simplified RF models. The combined method can be further applied to study other mechanical properties of complex fractured rocks and is particularly suitable for the cases with limited and scattered data from the fractured rocks in experimental and field investigations.