Comparison of machine learning and statistical approaches to estimate rock tensile strength

Abstract

Tensile strength is very important in drilling operations. The main objective of this study was to assess petrography, physical, and mechanical properties and predict the Brazilian tensile strength of sedimentary rocks by leveraging key parameters such as Schmidt hardness, compressional wave velocity, density, and porosity. A diverse array of predictive models was employed, encompassing simple regression, multivariate linear and nonlinear regression, backpropagation artificial neural network, gaussian process regression, classification and regression tree, K-nearest neighbor, random forest, and support vector regression. Based on thin section analysis and X-ray diffraction, the samples were identified. The sandstone samples were meticulously categorized into two distinct groups: arenite and litharenite. Additionally, the limestone samples were stratified into the categories of packstone to mudstone based on texture. The highest failure mode frequency of the samples under the Brazilian tensile strength test was identified as central fracturing. Upon meticulous examination, it was discerned that compressional wave velocity exerted the most substantial influence on Brazilian tensile strength estimates, while density exhibited the least impact. Comparing the outcomes derived from the diverse modeling techniques, it was unequivocally established that the support vector regression model showcased the highest level of performance for forecasting Brazilian tensile strength. This was evidenced by the remarkable coefficient of determination of 0.99 along with an impressively low root mean square error of 0.03.

Keywords

• Rock tensile strength
• Microscopic studies
• Machine learning and statistical techniques
• Physical and mechanical properties
• Sedimentary rocks