Utilizing Artificial Intelligence Techniques for Predicting Rock Failure Parameters

Abstract

Abstract Maintaining a stable borehole is one of the major tasks during drilling operations. During the drilling, borehole breakout and drilling induced fractures are the two main instability problems which may lead to stuck pipe, sidetracking, and loss of circulation. To evaluate the stability of a wellbore, a constitutive model is required to estimate the stresses around the wellbore coupled with a failure criterion to predict the ultimate strength of reservoir rocks. The Mohr-Coulomb failure criterion is one of the commonly accepted criteria for rock strength estimation at a given state of stress. This failure criterion is mainly contributed from the cohesion and coefficient of internal friction parameters, which are determined by laboratory measurements. The laboratory measurements, although more reliable, are expensive and time-consuming. This paper discusses artificial intelligence models particularly multilayer perceptron (MLP) and support-vector regression (SVR) for predicting cohesion and coefficient of internal friction from elemental spectroscopy and petrophysical properties. Elemental spectroscopy, density, porosity, cohesion, and coefficient of internal friction data presented in this paper are based on various geological formations. Cohesion and coefficient of internal friction are determined through a rock mechanical test in the laboratory, while elemental spectroscopy data were obtained from X-ray fluorescence (XRF) analysis. We divide the data set into training and testing data. Training data is used to train MLP and SVR then establishes the cohesion prediction models. Similarly, training data is used to train and construct the MLP and SVR-based coefficient of internal friction models. Both models are then examined using the testing data. Cohesion and coefficient of internal friction predicted from MLP and SVR match well with the laboratory measurements. Two quantitative measures for estimation accuracy are used including coefficient of determination and mean absolute percentage error. Cross-correlation plots of predicted cohesion and coefficient of internal friction and the experimental results show very good coefficient of determination and relatively small error. The results demonstrate that amongst the MLP and SVR models, the models whose inputs are grain density, porosity, and elemental spectroscopy are the best models. From a practical point of view, the application of artificial intelligence techniques as a new method for indirect estimation of rock failure parameters are beneficial especially when the amount of core samples are relatively few.