Improving estimation of rock mechanical properties using machine learning

Rock mechanical properties (e.g., uniaxial compressive strength or UCS, Young’s modulus, and Poisson’s ratio) are important input parameters for the geotechnical assessment and excavation designs. Two common methods to obtain these parameters are laboratory testing and geophysical logging. The former delivers probably the most reliable results, but can be costly and time-consuming and a lot times are challenging to source sufficient amount of samples. Alternative ways to better predict rock mechanical properties are needed. In this case study, we applied XGBoost machine learning algorithm to correlate laboratory and geophysical logging data with 3 mechanaical properties of UCS, Young’s modulus, and Poisson’s ratio. The proposed machine learning approach better predicted UCS values with a smaller mean absolute error (MAE) and root mean square error (RMSE) and a larger R2 . Similarly, better results were obtained for Young’s modulus prediction using the XGBoost. However, poor correlations existed between the inputs of geophysical and Poisson’s ratio, most likely due to the uncertainties associated with the acquisition of Poisson’s ratio data and the nature of this parameter. This study concluded that machine learning approach has the potential to predict rock mechanical properties more reliably than the conventional methods, and further study is underway to have more quantitative and detailed analysis with more data inputs and other machina learning models.