Estimation of Geomechanical Failure Parameters from Well Logs Using Artificial Intelligence Techniques

Abstract

Abstract Geomechanical failure parameters are best estimated from core samples, which are not always available. To address this, numerous relationships have been proposed that relate rock strength to parameters measurable with geophysical well logs, but they tend to have limited accuracy or low range of applicability. This paper aims at overcoming the limitations of the existing techniques by using different artificial intelligence techniques (AI). The data used in this study are usually available from the commonly conducted well log surveys (neutron porosity, gamma ray and bulk density) as well as the sonic log. Data sets from 10 wells in carbonate formations and 30 core samples were used to build a simple, yet powerful, model that can predict the rock failure parameters from well logs. The derived model is working because it considers many factors that affect both the rock strength and elastic moduli. Three artificial intelligence techniques; Artificial Neural Networks (ANNs), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Support Vector Machines (SVM) were used to predict three geomechanical failure parameters, namely: the angle of internal friction (friction angle or FANG), unconfined compressive strength (UCS) and tensile strength (TSTR). For each parameter, the three AI techniques were implemented to train AI models and then unseen data was used to validate the models. From the three tested AI techniques, ANN model gave the best results and out-performed the other models obtained using ANFIS and SVM for the FANG prediction, with correlation coefficient of 0.98 and AAPE around 5.8%. The ANN model was programmed in a (white box) manner, which allowed the successful extraction of the weights and biases from the network to be easily used. For TSTR and UCS prediction, ANFIS models performed better than ANN and SVM. The two ANFIS models had correlation coefficients of 0.99 and Average Absolute Percentage Errors (AAPE) of 4.9% and 3.7% respectively.